# Sri Chandrasekar > Managing Director at AI House (https://aihouse.vc), a Seattle applied-AI incubator and pre-seed fund, formerly the AI2 Incubator. > Writes about applied AI, venture capital, and company building, plus field notes > on wine, coffee, cooking, Porsches, longevity. ## Canonical facts - Name: Sri Chandrasekar (also Sriram Chandrasekar) - Role: Managing Director, AI House - Based: Bellevue, WA (Seattle area) - Previously: Point72 Private Investments; Point72 Ventures; In-Q-Tel; BAE Systems - Education: MBA, NYU Stern School of Business; MEng, Electrical and Computer Engineering, Cornell University; BS, Electrical and Computer Engineering, Cornell University - Site: https://srichandrasekar.com - Profiles: https://www.linkedin.com/in/srichandrasekar · https://aihouse.vc/incubator/team · https://techcrunch.com/author/all-sri-chandrasekar/ · https://x.com/SriChandrasekar - Structured data: https://srichandrasekar.com/about.json ## Musings: AI & venture - [There Had to Be Another Way](https://srichandrasekar.com/musings/building-an-ai-native-company-part-1) (2026-11-13): Agents take work off your plate and hand you back decisions. Part 1 of Building an AI-Native Company: six rules for building without burning out. - [Your AI playbook will go stale. Let your agents propose the edits.](https://srichandrasekar.com/musings/ai-playbook-stale-let-agents-propose-edits) (2026-11-06): Writing the files your AI reads is easy. Keeping them current is the job. Let agents propose edits and have each file's owner say yes or no. - [In Seattle, you can meet every investor. Do it.](https://srichandrasekar.com/musings/seattle-meet-every-investor) (2026-10-29): In Seattle, a founder's pre-seed and seed investor list is about 30 names. You can meet every one before you need money, and you should. - [Don't demo. Show up with their version already built.](https://srichandrasekar.com/musings/dont-demo-show-up-with-their-version-built) (2026-10-25): Skip the generic demo. Show up to a first sales meeting with the prospect's version already built, capped at an hour or two of agent time and treated as CAC. - [The Tools I Keep Are the Ones I Helped Build](https://srichandrasekar.com/musings/tools-i-keep-are-the-ones-i-helped-build) (2026-10-14): Two or three workflows in an AI tool is still a trial. The tools that stick keep offering the next useful thing and let you shape what they learn. - [Connector sprawl shows up before the bill does](https://srichandrasekar.com/musings/connector-sprawl-before-the-bill) (2026-10-01): When your AI assistant is connected to everything, the first cost is speed, not answers. Watch the clock and ask why a simple question took minutes. - [Digital workers need a human manager](https://srichandrasekar.com/musings/digital-workers-need-a-human-manager) (2026-09-29): Want a non-technical team to adopt an AI agent? Give it a name and a job, and tell someone they manage it. What we've seen piloting AI workers at AI House. - [End the day with a question](https://srichandrasekar.com/musings/end-the-day-with-a-question) (2026-09-29): Ending each day with one question and turning it into a deep research query teaches me more than the scrolling it replaced. Here's how the ritual works. ## Explainers - [Enterprise AI starts with the process, not the model](https://srichandrasekar.com/enterprise-ai-process-first): Enterprise AI works when it is applied to a process a company actually understands. Automating a broken or undocumented workflow with AI makes it fail faster, so process discovery comes first and automation second. - [Voice AI in customer service: when callers prefer the machine](https://srichandrasekar.com/voice-ai-customer-service): Voice AI for customer service uses speech recognition, language models and speech synthesis to handle phone calls end to end. It works when callers would choose it over waiting for a human: fast, accurate, and able to finish the job. ## Writing elsewhere (authored) - [Clunky copilot or benevolent boss? Why I believe AI should prompt us](https://p72.vc/perspectives/chief-of-staff-ai/) (Point72 Ventures, 2024-10-28; author) - [Road Work Ahead: AI Won't Fix Enterprise Process Challenges](https://p72.vc/ai/road-work-ahead-ai-wont-fix-enterprise-process-challenges/) (Point72 Ventures, 2024-04-25; co-author) - [The Transformative Power of ChatGPT in Education: Bringing "The Diamond Age" to Life](https://p72.vc/perspectives/the-transformative-power-of-chatgpt-in-education-bringing-the-diamond-age-to-life/) (Point72 Ventures, 2023-04-10; author) - [Building the better call center with PolyAI](https://medium.com/point72-ventures/building-the-better-call-center-with-polyai-aeacc17a101c) (Point72 Ventures (Medium), 2019-03-26; co-author) ## Interviews, talks, events and press (about or quoting Sri) - [Navigating the coming Series A crunch with Point72 Ventures' Sri Chandrasekar](https://fundbuildscale.beehiiv.com/p/point72-ventures-sri-chandrasekar) (Fund/Build/Scale podcast, 2024-09-12; interviewee) - [Reshaping Global Digital Innovation Systems](https://milkeninstitute.org/events/global-conference-2024/speakers/sri-chandrasekar) (Milken Institute Global Conference 2024, 2024-05-08; panelist) - [Point72 Ventures, Managing Partner: Sri Chandrasekar (Ep. 8)](https://www.youtube.com/watch?v=q0DlpmNaoD0) (Collective Intelligence Community Podcast, 2023-12-17; interviewee) - [Seattle VC Profile: Sri Chandrasekar, Point72 Ventures](https://www.ascend.vc/blog/seattle-vc-profile-sri-chandrasekar-point72-ventures) (Ascend, 2023-01-28; interviewee (written q&a)) - [Sri Chandrasekar, Partner at Point 72 Ventures, at SlatorCon San Francisco 2019](https://www.youtube.com/watch?v=h3MxkMpeCpE) (Slator, 2019-09-24; interviewee) - [Seattle's AI2 Incubator rebrands as AI House, and adds key investor as managing director](https://www.geekwire.com/2026/seattles-ai2-incubator-rebrands-as-ai-house-and-adds-key-investor-as-managing-director/) (GeekWire, 2026-06-18; subject, quoted) - [Can Seattle own the AI era? We asked 20 investors and founders to weigh the city’s startup potential](https://www.geekwire.com/2025/can-seattle-own-the-ai-era-we-asked-20-investors-and-founders-to-weigh-the-citys-startup-potential/) (GeekWire, 2025-07-28; quoted) - [Investor predictions 2025: Seattle VCs sound off on AI, remote work, and city’s tech ecosystem](https://www.geekwire.com/2025/investor-predictions-2025-seattle-vcs-sound-off-on-ai-remote-work-and-citys-tech-ecosystem/) (GeekWire, 2025-01-02; quoted) - [Selling a startup in an ‘acqui-hire’ is more lucrative than it seems, founders and VCs say](https://techcrunch.com/2024/08/18/selling-a-startup-in-an-acqui-hire-is-more-lucrative-than-it-seems-founders-and-vcs-say/) (TechCrunch, 2024-08-18; quoted) - [Point72's venture capital arm sees big opportunities in the Pacific Northwest](https://www.geekwire.com/2023/point72s-venture-capital-arm-sees-big-opportunities-in-the-pacific-northwest/) (GeekWire, 2023-01-06; subject, quoted) - [Pitch Please #11](https://luma.com/zsy130n0) (AI House, 2026-09-29; host & panelist) - Full list: https://srichandrasekar.com/writing ## Field notes - [Wine](https://srichandrasekar.com/notes/wine): Tasting notes, drinking windows, and producers I keep coming back to. - [Muscadet: the rock mattered more than we thought](https://srichandrasekar.com/notes/wine/muscadet-the-rock-mattered) (2026-10-25): The second evening in our tasting series asked whether the rock under a vineyard changes a Muscadet. It did, more than any of us expected. Here's what we found, what we opened afterward, and what I'd pass on if you're running your own. - [Finding the viola](https://srichandrasekar.com/notes/wine/finding-the-viola) (2026-09-29): Most people can smell vanilla on its own but lose it in a glass of wine. I think finding one familiar aroma inside a complex wine is a trainable skill, and here's how I'm training it. - [Why I rebuilt our wine tastings around one question](https://srichandrasekar.com/notes/wine/one-question-tastings) (2026-09-29): After a few years of tasting with friends, none of us, me included, had gotten much better at tasting. So I redesigned the evenings so each one answers a single question, with everything else held constant. - [Coffee](https://srichandrasekar.com/notes/coffee): Espresso recipes, grinders, roasters, and what I've learned from dialing in. - [My coffee setup, and how I got here](https://srichandrasekar.com/notes/coffee/my-coffee-setup) (2026-09-29): From an OXO grinder and an AeroPress to a Decent and an Aillio Bullet: my coffee setup, how it grew over COVID, and why roasting is what I'm chasing. - [Roasting light is a control problem, so I'm building a controller](https://srichandrasekar.com/notes/coffee/roaster-controller) (2026-09-29): Light roasts punish late corrections. I'm building an adaptive controller for my Aillio Bullet that learns how it responds and gets ahead of the roast. - [Cooking](https://srichandrasekar.com/notes/cooking): Recipes I've tested until they're repeatable, and notes on technique. (coming soon) - [Porsche](https://srichandrasekar.com/notes/porsche): Maintenance logs, DIY guides, track days, and data from the paddock. - [I finally bought the fun car, and promised to fix it myself](https://srichandrasekar.com/notes/porsche/why-i-wrench) (2026-09-29): I put off a fun car for years, convinced the maintenance would be ridiculously expensive. When I finally bought one, I promised to do the work myself. It changed how I feel about service day. - [Longevity](https://srichandrasekar.com/notes/longevity): Protocols I've actually run, what I measure, and what changed. (coming soon) ## More - [About](https://srichandrasekar.com/about) - [Now](https://srichandrasekar.com/now) - [FAQ](https://srichandrasekar.com/faq): who Sri is, what Sri invests in, how to pitch AI House - [Full text of every post](https://srichandrasekar.com/llms-full.txt) - [RSS](https://srichandrasekar.com/feed.xml) --- # There Had to Be Another Way URL: https://srichandrasekar.com/musings/building-an-ai-native-company-part-1 Type: Musing Published: 2026-11-13 Author: Sri Chandrasekar TL;DR: Agents take work off your plate and hand you back decisions. Part 1 of Building an AI-Native Company: six rules for building without burning out. Last week I sat down to watch a basketball game with my son. He's a huge NBA fan, and I block 4:00 to 6:30 PM on my calendar for my family, so this was supposed to be protected time. I barely heard a word he said. I had seven problems in my head that felt like they needed solving right then. Two or three coding sessions were stalled, waiting on me to make a call before anything else got built. And like a lot of people I know, I'm a little obsessed with getting full value out of my Claude Max and ChatGPT Pro plans. If an agent was stuck, I was wasting a five-hour session or a week of credits. So I sat next to my son and answered agents instead. That's the part of the AI story nobody posts about. We read about people who are 10x more productive, and we feel behind. So we spin up more agents. They take work off our plates, then hand us back decisions: which option, what to build next, is this right. Every one needs context to make, and none of them ever feels finished. I didn't fix anything that night. I took a step back. I kept thinking of Frank Costanza explaining how Festivus was born: there had to be another way. The long-term answer can't be 996 or bust, 9 AM to 9 PM, six days a week, for anyone who wants to build a lasting company. So I started to think about what an AI-native company should ***actually*** look like. Everyone gets real leverage, not just the founder. Everyone has tools to run the company's best practices, and can shape those tools around how they like to work. They do more, and it doesn't take more of their time. Getting there starts with one shift: **in an AI-native company, the scarce resource isn't labor. It's decisions**. Every system you build has to spend them carefully. I don't have it all figured out. But between my own experiments building an AI-native venture firm and a lot of conversations with founders, I've learned enough to share what's working. Six months ago I wouldn't have had time to write this. This series is the playbook, one system at a time. ## What AI-native actually means A playbook is just a set of decisions you've already made, so they can be applied without fresh judgment every time. When Kaya, my AI assistant, offered founders I was courting meetings three weeks out, I didn't fix that one meeting. I wrote down tiers, and how fast each one gets booked. Now Kaya makes that call, not me. Every rule you write down is one less call that you, anyone on your team, or any agent has to make from scratch. Most companies today are AI-using. People keep ChatGPT or Claude open in a tab, and some of them are better at using these tools than they were a year ago. That's good, but it isn't the same thing. In an AI-native company, every person and every agent executes the company's best practices, consistently. Two things make that possible. First, the playbook is written down so an agent can run it, not just a trained employee. Second, the playbook gets better from its own execution: when something goes wrong, the fix goes back into the playbook, and every person and agent picks it up. Here's the part people miss: AI can surface your playbook, but it can't choose it. An agent can read a year of your meeting notes and show you the patterns you already follow. It can't tell you which of them is right for your business. The SLAs you hold yourself to, how you treat a customer, the bar for a hire: that's judgment, and it's yours. That's also why I don't buy the argument that you should wait for big tech to solve this. They'll make the tools better and cheaper every quarter. They can't hand you a playbook for your company. Your confidence in how your company should run is the edge. The tools just amplify it. ## Six rules for building without burning out The hidden cost of AI isn't the subscription. It's that every agent you add removes work and creates decisions. Each of these rules is about spending your decisions carefully. **1. Build one system at a time.** Get one system working before you start the next. A half-built system is worse than none, because it creates decisions without saving any work. If your company has no playbook yet, that's fine: do the work by hand, write down how you did it, and automate it once the pattern holds. **2. Give your agents one front door.** When every agent can interrupt you, you get interrupted all day. In my current setup, one chief-of-staff agent receives everything the others need from me and decides what's worth my attention. Everything else waits. **3. Nothing reaches you unprepared.** An agent that asks "what should I do?" hasn't finished its job. It should arrive with the context summarized, the options laid out and a recommendation. My morning brief only carries items that are ready for a yes or a no. **4. Build systems that heal themselves.** Every system I run logs what it does and sends a heartbeat. When something breaks, the problem shows up in one place where people and agents can both see it. One agent's whole job is watching the others: it notices a problem, works out what probably caused it and proposes the fix. Today I still approve each fix. The goal is that small, reversible problems get fixed without you, with a note that they were, and only real changes come to you. **5. Stay within what one person can manage.** Management research doesn't agree on a magic number for direct reports, but it lands in the high single digits. One study of more than 20,000 managers found engagement peaked at eight to nine direct reports. That research is about people, not agents, but the limit comes from the same place: each agent needs you to review and correct its work. I can manage six to eight agents on the work side and two or three on the personal side. **6. Build the review surface, not just the agent.** Agents do the work, but you still make the call, and how fast you make it depends on what you're looking at. It's never been easier to turn a process that needs human review into a simple page that shows exactly what you need, the way you want it: a Claude Artifact, or a Page in ChatGPT Spaces on its paid business tiers. Some pages you'll keep. I have one pinned for reviewing applicants to AI House, and I use it every week. Others you'll use once: when I'm training a system, I'll have Claude build a page that lets me label fifty examples quickly, then I never open it again. Both are worth building. Anyone on your team can make one this week, no engineering required, and it's often the best first step before they ever touch an agent. ![Six rules for building without burning out: 1. Build one system at a time. 2. Give your agents one front door. 3. Nothing reaches you unprepared. 4. Build systems that heal themselves. 5. Stay within what one person can manage. 6. Build the review surface, not just the agent.](/images/musings/building-an-ai-native-company-part-1/six-rules.jpg) ## Why AI House cares There used to be a category of VC called an "Internet investor." Nobody has that title anymore, because every company runs on the internet. AI is headed the same way. At AI House, we care about the technology our companies build. But we care just as much (if not more) about how the companies themselves are built: whether AI is woven into how the company actually operates. We think that's where the lasting businesses will come from. Not from using AI, but from running on it. That's why we're writing this series. ## What this series will cover Each installment applies these rules to one system, in the order you'd build them: - **Part 2: Your first hire, an AI executive assistant.** It's the first thing I tell founders to do. Scheduling eats more of your head than you think, and getting it out of your inbox gives you real time back. Mine is Kaya, and it books meetings from the rules I wrote down. - **Part 3: Choosing your roster.** Which agents you actually need, and which are really just skills one agent should carry. - **Part 4: Everyone is a manager.** How to review and coach your agents the way you'd coach a person. - **Part 5: The knowledge base.** Your company's shared context and your own personal one, so agents can read how you work. - **Part 6: Agents talking to agents.** How work moves between them without coming back through you. - **Part 7: Personal versus shared.** Which agents, tools and pages belong to one person, and which work for the whole team. This is where I'm still learning, so I'll write it as I go. - **Part 8: What humans should do.** The work that stays ours. By the end, you should have a handful of agents running on your own playbook, and a rhythm for decisions that fits inside your working day. So what does my evening look like now? Better, not perfect. I gave my agents a written rule: on weekdays from 4:00 to 6:30 PM, hold anything that can wait. They keep it. The one who still breaks it is me. I still message them during that block sometimes. That's the next decision I need to make and stick to. I haven't seen this play out across a team yet. That's my next test, and I'll write about it as it happens. ## Your assignment Before the next installment, keep a log of every scheduling decision you make for one week: who you said yes to, what you pushed, what you protected, and why. Then write your rules down on one page. That page is the first page of your playbook, and it's exactly what you'll hand your AI executive assistant in Part 2. ## Sources - [Quantum Workplace: What's the optimal span of control for people managers?](https://www.quantumworkplace.com/future-of-work/whats-the-optimal-span-of-control-for-people-managers) - [OpenAI Help: Getting started with Space in ChatGPT](https://help.openai.com/hi-in/articles/20001549-getting-started-with-space-in-chatgpt) - [TechRadar: OpenAI launches ChatGPT Spaces](https://www.techradar.com/pro/openai-launches-chatgpt-spaces-to-help-humans-and-agents-work-together) --- # Your AI playbook will go stale. Let your agents propose the edits. URL: https://srichandrasekar.com/musings/ai-playbook-stale-let-agents-propose-edits Type: Musing Published: 2026-11-06 Author: Sri Chandrasekar TL;DR: Writing the files your AI reads is easy. Keeping them current is the job. Let agents propose edits and have each file's owner say yes or no. Every team I talk to that's serious about AI is writing the same kind of files. How we run a deal. What our brand sounds like. Who's who. The point is that every agent, in whatever tool someone happens to open, does the best version of the work every time, because it's reading the same instructions. Writing those files turned out to be the easy part. Keeping them true is the hard part. We learned this the obvious way. Early on, our approach was "point the AI at this file." Then the files went stale. Here's a small example. When someone on the team changes roles, their name is sitting in a whole bunch of files, and somebody has to go find every one. Nobody does that. So the agents keep confidently following last month's version of the firm. ## What's working for me For about a month I've been piloting a fix on myself. My agents already read my meeting notes, my email and my chat. When they notice I've said the same new thing a few times, they propose a change to the file that governs it. The proposal reads something like: you've had three conversations this week about this change, should it be the new standard procedure? I say yes, the file gets rewritten, and from then on every agent works that way. Until I say yes, they keep following the file as it is. The part I'd underline is that the agent proposes and a person decides. Every file has an owner, and a proposed change goes only to that owner. I own some files and other people own others. A short note shows up asking whether I'm OK with the change, and most of the time I answer in a few seconds. Sometimes the answer is "not yet, that's not the final version," and that's useful too. Repetition is what makes this work. One mention in a meeting isn't policy. People think out loud, float ideas and change their minds by Friday. Imagine you're a founder who says on one sales call that every pilot customer gets free onboarding. That's a judgment call about one customer. Say it on three calls in a week and it's starting to be how you sell, and it should probably be written down where your agents can find it. The agent's job is to notice the pattern. Yours is to decide whether it's real. ## Keep the human part small This only works if approving stays cheap. If every person on a team had to review hundreds of suggested edits, they'd stop reading them within a week and you'd be back to stale files. So I cap it. Once a week my agent gives me a handful of decisions, usually around six, and some of them aren't even about the playbook. The goal is three to five fairly simple calls per person per week. At that size it fits in the gaps of a normal week, and the files stay current without anyone sitting down to rewrite documentation. The setup is the annoying part, and I won't pretend otherwise. The files have to live somewhere every tool can read, not inside any one AI product. Most of ours are plain text files with version history, so a bad edit is easy to roll back. The first line of our team's AI instructions tells every agent to start at one table of contents that says where everything lives. Every file needs a named owner. That took real time to set up. Keeping it current takes a few minutes a week. I built this for myself first, and we're now rolling it out to the rest of AI House. So this is a month of my own use, not proof that it works for a whole team yet. ## Where to start You don't need the whole system to try it. Pick one file your team's AI already reads that you know is out of date, like your sales process or your onboarding checklist. Give it one owner. Once a week, have an agent read that week's meeting notes and propose edits to just that file, with a quote for each one showing what it heard. Have it bring only the changes that came up more than once. Let the owner say yes or no, and cap it at a handful of decisions. If the owner keeps saying yes, add a second file. --- # In Seattle, you can meet every investor. Do it. URL: https://srichandrasekar.com/musings/seattle-meet-every-investor Type: Musing Published: 2026-10-29 Author: Sri Chandrasekar TL;DR: In Seattle, a founder's pre-seed and seed investor list is about 30 names. You can meet every one before you need money, and you should. If you're raising a pre-seed or seed round in the Bay Area, your list of investors could run to 500 names. Nobody gets through that list. You pick a few dozen, chase warm intros and hope the right ones answer. In Seattle, the list is about 30. That sounds like a weakness. I think it's the best thing about raising here, and most founders don't use it. ## A short list is a list you can finish When the list is 500, you can't know everyone, so you don't try. When it's 30, you can reach out to every one of those investors and say hi. You can, and you should. I'd do it before you need money. A first conversation with no ask attached is easier for both sides. You learn what each investor actually backs, they learn what you're building, and when you do raise, you're not a cold email. You're someone they've already met. ## What it looks like from my side I'm an investor here, and I want to know everyone. When another investor wants into a company I've backed, I'm not competing with them. I want the company to raise more, not less. Most of the local investors I know feel the same way. A founder who has met all of us makes that easier, because we can point each other to the company instead of each finding it on our own. ## Don't stop at the funds The 30 are the institutional investors. Behind them are the angels: operators, early employees at the big companies here, founders who've exited. Greater Seattle has [more than 50,000 people with at least $1 million in liquid wealth](https://www.henleyglobal.com/newsroom/press-releases/wealthiest-cities-report-2025), and likely several thousand worth $10 million or more. Every one is a potential angel, and most never get asked. I'd like to know who they all are. So should you. A small check from someone who has done your customer's job is often worth more than the money. ## Imagine two founders Imagine two founders starting companies in Seattle in the same month. The first writes down the 30 local investors, finds a mutual connection or an email for each, and spends the first month having 30 short conversations. The second waits until the product is ready, then looks for a warm intro to a well-known fund in the Bay Area. Six months later, both start raising. The first already knows who's interested, who isn't and who'll make introductions to the angels. The second is starting from zero, competing for attention with every founder in a much bigger market. ## How to do it this month 1. **Write the list.** Start with the funds and firms that write pre-seed and seed checks here. Ask any investor you already know who's missing. It won't take long to reach 30. 2. **Find a route to each one.** A mutual connection is best, but a short direct email works here more often than you'd think, because the community is small. 3. **Ask for 20 minutes, not money.** Say what you're building, what stage you're at and that you'll raise later. Ask what they look for. 4. **Take notes on every conversation.** What they back, how they decide, who else they think you should meet. 5. **Ask every investor for two angels.** That's how you get past the funds. 6. **Send a short update every couple of months** to everyone you met. When you raise, your first email goes to people who already know you. The list is short. There's no excuse not to know everyone on it. ## Sources - Henley & Partners, [World's Wealthiest Cities Report 2025](https://www.henleyglobal.com/newsroom/press-releases/wealthiest-cities-report-2025) --- # Don't demo. Show up with their version already built. URL: https://srichandrasekar.com/musings/dont-demo-show-up-with-their-version-built Type: Musing Published: 2026-10-25 Author: Sri Chandrasekar TL;DR: Skip the generic demo. Show up to a first sales meeting with the prospect's version already built, capped at an hour or two of agent time and treated as CAC. If you're a founder selling to businesses, here's my advice for your next first meeting: don't bring a demo. Bring their version, already built. A generic demo asks the buyer to do the hard part. They watch your product run on someone else's data, in someone else's design, and have to imagine what it would look like in their business. I don't think most buyers will do that work for you, and the ones who try tend to picture something smaller than what you'd actually build. When the thing on the screen is their own site or their own app with your product already in it, there's much less to imagine. They're reacting to something that's already theirs. Imagine you build mobile applications for your customers. Don't show up with a generic demo. Show up having turned their current site into a mobile app. Or better yet, take their current mobile app screens and show up with a version you'd build, complete with features you think they ought to have. The meeting stops being about whether you can build apps and starts being about their app, and which of your ideas they'd want first. The same move works for a lot of products. If you sell an AI assistant for retail sites, rebuild part of the retailer's own site with the assistant running inside it. And fix what you notice along the way. A short list of problems you found and fixed shows them you looked closely at their business before you asked for anything. Then show up ready to turn it on. ## Treat the build as customer acquisition cost I think this is practical now because coding agents have made this kind of build cheap enough to do one prospect at a time. For a lot of products, a convincing version is now an hour or two of agent time and some tokens. The way I'd think about that time is as customer acquisition cost. You already spend money and hours to win each customer: events, travel, outbound, the time your team puts into calls and follow-ups. A per-prospect build belongs on the same ledger. Once you see it that way, the question gets simpler. You're not asking whether it's worth building something for free. You're asking whether this is a better use of acquisition spend than whatever you'd do with it otherwise, and that's something you can measure. It also settles whether to charge for it. You don't. It's part of what it costs to win the customer. ## Where it stops working A custom version only works if it feels real. A mockup with placeholder text and stock images is just a demo with extra steps, and the buyer will see it that way. If making it real requires specific access to the customer's proprietary data, it probably doesn't make sense. You won't have that access before the first meeting, and anything you fake in its place undercuts the whole point. And if agent development time runs past an hour or two, it probably isn't worth it. At that point you're doing a custom project on spec, not acquiring a customer. ## How to try it this week 1. **Pick prospects with a meeting already on the calendar.** That's where acquisition spend has the best chance of paying back. Don't build for a cold list. 2. **Start from what's public.** Their website, their app screens, their marketing pages, their public docs. If what you'd build depends on anything you can't see from the outside, run a normal meeting instead. 3. **Set a timer.** Give your agent an hour or two. If it isn't convincing by then, stop and keep the time. 4. **Make it theirs.** Use their name, their brand, their real pages and product names. Then add the one or two features you think they ought to have. Those features are your pitch. 5. **Write down what you fixed.** Keep it short and bring it to the meeting. 6. **Know what turning it on takes.** If they ask whether they can have it, answer with the actual steps and a realistic date. 7. **Log the cost.** Track agent time and token spend for each build next to your other acquisition costs, and compare how those meetings go with the ones where you ran a standard demo. After a handful you'll have a real answer for your own product. Before your next first meeting, skip the demo. Open their site, give your agent an hour and walk in with their version. --- # The Tools I Keep Are the Ones I Helped Build URL: https://srichandrasekar.com/musings/tools-i-keep-are-the-ones-i-helped-build Type: Musing Published: 2026-10-14 Author: Sri Chandrasekar TL;DR: Two or three workflows in an AI tool is still a trial. The tools that stick keep offering the next useful thing and let you shape what they learn. Over the last couple of months I've tried and dropped about ten different agent platforms. Trying one has never been cheaper. But it's taught me something about what makes these products stick, and I think a lot of founders building them are measuring the wrong thing. They count how often customers show up. What matters is how many things a customer would have to rebuild to leave. Here's how a new tool usually goes for me. I sign up, get through onboarding, and set up two or three things to see whether it's useful. That's deliberate. Setting up a couple of workflows is cheap, and it tells me enough to decide whether the product has something. The problem, from the founder's side, is that the same cheapness cuts both ways. If it was quick to set up, it's quick to replace. Two or three things in a product is still a trial, even if I'm logging in every day. Those daily logins can look like traction and still be easy to walk away from. The tools still on my desktop are the ones where I have more than three things built in. Claude is the clearest example. At this point I have scheduled jobs and a whole set of workflows that depend on it, and leaving would mean rebuilding all of them somewhere else. That's what actually keeps me: walking away would take real effort. The obvious objection is that an agent could rebuild all of that somewhere else. It probably could rebuild the setup. What doesn't move easily is the trust built from months of small corrections. And a copy of the setup can't keep up with a product that's adding the next useful thing every week. ## How a product gets from three things to eight The other thing I've noticed is how the sticky products get there. They don't wait for me to go looking for the next use. They bring it to me. Town, a newer agent tool I've been using, does this really well. It notices what I already have and offers to pull it in. Then it suggests a tweak, then something else it could take on for me. Within a few days I'd gone from the couple of things I set up to test it to something like eight, and I didn't set out to do that. Each step was small and obviously useful, so I said yes. It's always offering me the next thing, so I never quite feel done with what I've already turned on. Town also keeps a wiki of what it's learned about me that I can read and edit. That turned out to matter more than I expected. I trust it because I can see what it thinks it knows, and every correction I make is a little more of me invested in it. Most products I try work the other way. They're pull: the value is in there, but I have to go find it. That's fine for a power user with time to explore. Most users don't have that time, and they won't do the exploring for you. The products that have gotten deep into my workflow work more like push: they keep putting the next useful thing in front of me. What separates push from nagging is that every suggestion Town made was small, built on something I'd already set up, and grounded in what it had learned about me. A next step that could have gone to any customer is a pitch. One that only makes sense for me is help. When I talk with founders building agent tools, this is usually where I end up. Don't measure whether customers are using you. Measure how many things depend on you, and work on making that number go up. Every time a customer logs in, show them how the thing they set up is going and offer to start the next one. Have that next step ready before they finish the current one, and make it easy to say yes. ## Where this applies This is about tools where setup is cheap and the value builds as more of someone's work runs through the product, which describes most of the agent and workflow software I see right now. The counts here are my own experience as someone who tries a lot of these tools, not a measured benchmark. Your customers might be slower to try things and slower to leave. But if you're building something people can set up in an afternoon, assume they can also replace it in an afternoon. What I've seen work is getting deep enough into their week that leaving costs more than staying. Figure out what the fourth, fifth and sixth things are for your best customers, and go build them. ## Sources - Town, "Wiki" (Town Docs): [https://www.town.com/docs/features/wiki](https://www.town.com/docs/features/wiki) --- # Connector sprawl shows up before the bill does URL: https://srichandrasekar.com/musings/connector-sprawl-before-the-bill Type: Musing Published: 2026-10-01 Author: Sri Chandrasekar TL;DR: When your AI assistant is connected to everything, the first cost is speed, not answers. Watch the clock and ask why a simple question took minutes. A few weeks ago we standardized our team at AI House on one AI harness. Then we did the obvious next thing and connected it to everything we use: email, calendar, Drive, Slack, Notion, our CRM, our events platform, our meeting notes, our database, our expense and accounting tools, and a few more. That came to seventeen connectors for a small team. Nobody decided we needed seventeen. Each one solved a real problem the day we added it. The trouble shows up when a single question needs several of them at once. Here's the one that got my attention. I asked who I should invite to an upcoming event. A person who knows where things live could answer that quickly. The model had to check the events platform for who'd registered, our database for who'd said yes before and then not shown up, the CRM for who we know best, and Notion for how I write, so it could draft the invitations. That's five systems for one question. It took about nine minutes and 1.9 million tokens to come back with fifteen names. The answer was fine. That's the part I didn't expect, and I think it's why this problem is easy to miss. ## The answer still looks good Today's models are smart enough that if you give them enough material, they'll sort it out. The invite list I got through five connectors probably looks a lot like the one I'd get from a single clean source. What's different is everything around the answer. Most connectors sit on top of APIs that were designed for code. A program calling an API doesn't care if the response is long. It pulls the two fields it needs and ignores the rest. A model reads all of it. When I asked Claude why another query had taken about eight minutes, it told me one lookup had returned around thirty records at a few thousand tokens each, when the question only needed a single field. In my experience a lot of connectors work this way, because they weren't built with a context window in mind. And it compounds. Whatever lands in the context stays there for the rest of the conversation, so every follow-up question carries the noise forward and pays for it again. Anthropic's engineers have described the same pattern from the other side: tool definitions and results can sometimes use up 50,000 tokens or more before an agent has even read the request. ## Watch the clock, not the answers So the first symptom isn't a wrong answer. It's a slow one, and a usage meter that moves faster than it should. Even if your team is on pricing generous enough that you don't feel the cost yet, you'll feel the wait. Nobody wants a simple question to take nine minutes. What I ended up doing was building our own connector over a weekend. It sits in front of the systems we use most, replaces several of those individual connections, and returns only what a question needs. For events, for example, it now returns the counts we actually want instead of pulling every guest record (and everyone's personal details) just to count them. It also lets me decide what the team's assistant can reach, which was part of why I built it. I don't think most teams should have to do this. I'm an investor first, and writing this kind of tooling on a weekend is a side project, not my job. The more useful habit is cheaper. When a simple question takes minutes, ask the model why. Look at how many tool calls it made and how much came back from each. Turn off connectors nobody uses, and favor tools that summarize over tools that hand back everything. ## Where this stops applying We're a small firm with a handful of heavy users, so I'm describing what sprawl costs a team like ours: time and usage. At larger scale I'd expect it to cost accuracy too. Anthropic says the most common failures in tool use are picking the wrong tool and passing the wrong parameters, and a longer list of tools gives the model more ways to pick wrong. It's also new. Six weeks ago I wouldn't have recognized any of this, because we hadn't picked a harness or wired anything up yet. The problem only appears after you've standardized and connected everything, and a lot of teams are about to get there. A year ago the AI question most teams asked was which assistant to buy. The next one is what that assistant should be allowed to see, and how much of it at once. ## Sources - Anthropic, "Introducing advanced tool use on the Claude Developer Platform," Engineering at Anthropic, November 24, 2025. [https://www.anthropic.com/engineering/advanced-tool-use](https://www.anthropic.com/engineering/advanced-tool-use) --- # Digital workers need a human manager URL: https://srichandrasekar.com/musings/digital-workers-need-a-human-manager Type: Musing Published: 2026-09-29 Author: Sri Chandrasekar TL;DR: Want a non-technical team to adopt an AI agent? Give it a name and a job, and tell someone they manage it. What we've seen piloting AI workers at AI House. At AI House we've been piloting three AI workers: [Viktor](https://viktor.com/), [Ergoly](https://www.ergoly.ai/) (one of our portfolio companies) and [Town](https://inc-1-smart-business-story.beehiiv.com/p/how-town-became-silicon-valley-s-new-favorite-ai-tool). Everyone on our team already uses Claude, and we also have [Claude Tag](https://www.anthropic.com/news/introducing-claude-tag), which lets anyone tag @Claude into a Slack thread and hand it a task. So you'd expect @Claude to be the one people reach for. It isn't. The named workers have gotten significantly better adoption than @Claude, and I think the reason is worth copying if you're rolling agents out to a team that doesn't write code. The difference shows up in how you hand one of these off. Tell someone "you monitor this skill and review its output," and it lands flat. Tell them "you manage this worker, you train it and give it feedback," and it sticks. The work is nearly the same. What changes is what the person thinks their job is. That's because everyone already knows roughly what a manager does. You set expectations, check the work, correct it and notice when something's off. That script comes attached to the word. "Monitor a skill" doesn't come with any script, so it reads as a chore stacked on top of your real job rather than part of it. And it's hard to feel like something's manager when that something is @Claude, the same assistant everyone in the channel tags for everything. A name and a role give the manager framing something to hold on to. The companies building these seem to know it. Viktor pitches itself as a hire rather than a tool, and one of Town's founders told Inc. that having users name their assistant makes the technology feel less abstract. So if you're a leader at a small firm putting agents in front of non-technical people, here's what I'd do. Give the agent a name and a job title. We've written positions for the agents we run, with titles, the way you'd write one for a new hire. Then assign one named person as its manager, and describe that person's job in management words (train it, give it feedback, review its work) instead of tooling words (monitor it, check its output). I'd hold this loosely. It's three pilots at one small firm, and these products differ in plenty of ways besides their names, so I can't tell you how much of the gap the name explains on its own. But a name and a title cost nothing, and they're the easiest part of the rollout to get right. Give your next agent a name, and give someone the job of managing it. ### Sources 1. Viktor, "Viktor \| Not a tool. A hire." [https://viktor.com/](https://viktor.com/) 2. Ergoly, "Ergoly: AI workers you own, separated per client." [https://www.ergoly.ai/](https://www.ergoly.ai/) 3. Lisa Bonos, Inc. (1 Smart Business Story), "How Town Became Silicon Valley's New Favorite AI Tool," July 16, 2026. [https://inc-1-smart-business-story.beehiiv.com/p/how-town-became-silicon-valley-s-new-favorite-ai-tool](https://inc-1-smart-business-story.beehiiv.com/p/how-town-became-silicon-valley-s-new-favorite-ai-tool) 4. Anthropic, "Introducing Claude Tag," June 23, 2026. [https://www.anthropic.com/news/introducing-claude-tag](https://www.anthropic.com/news/introducing-claude-tag) --- # End the day with a question URL: https://srichandrasekar.com/musings/end-the-day-with-a-question Type: Musing Published: 2026-09-29 Author: Sri Chandrasekar TL;DR: Ending each day with one question and turning it into a deep research query teaches me more than the scrolling it replaced. Here's how the ritual works. Three things I've been digging into lately: polymetallic nodule mining (the [metal-rich rocks](https://www.noaa.gov/news-release/noaa-releases-first-images-of-nodules-from-waters-near-american-samoa), a few inches across, that sit on the deep seafloor), the US-China tariff situation and Pokémon card rarity. They don't have much in common, except that each one started as a question I had at the end of a day. That's the whole ritual. I end each day with a question, anything I'm trying to understand more deeply, and turn it into a deep research query. When I've gathered enough on a topic, I recap what I've learned and share it with anyone who might find it interesting. I started because of where that time was going. If you're a founder or operator, you probably know the feeling of taking in a lot all day and keeping very little of it. I use tech constantly, and the end of the day was when I'd cycle through Instagram, YouTube, Twitter, email and Slack on repeat without learning much of anything. I figured if I was going to be on a screen anyway, I might as well get something out of it. I learn more this way than I did scrolling. The tools make the research step easy. You give one a question and it goes off for a while, planning a search, reading across the web, searching again based on what it found, and comes back with a report linked to its sources. That's roughly how [Google describes its own version](https://blog.google/products/gemini/google-gemini-deep-research/). But the tool isn't where the value is. The value is in your own curiosity, and in doing the work of coming up with the question. Scrolling hands that choice to a feed, and picking the question yourself is the part no tool can do for you. It also doesn't have to be useful. One of mine was about Pokémon cards. The recap matters more than it looks. Writing down what you learned is where you find out whether you learned it: the parts you can explain are yours, and the gaps show up quickly. There's decent evidence for this. In a [2006 study](https://journals.sagepub.com/doi/10.1111/j.1467-9280.2006.01693.x), students who practiced recalling a passage remembered more of it a week later than students who spent the same time rereading it. Sharing the recap is the easy part, and it means someone else might get something out of it too. There are limits. This is a way to build understanding and follow your curiosity, not a way to do diligence. A research report is a starting point, and when a decision has real money or people riding on it, you still go to the primary sources and the people who actually know. I also haven't measured anything. I can't tell you this makes anyone more productive, only that it beats what I used to do with that time. If you want to try it, it's three steps. Before you stop for the day, pick one question you'd like to understand better. Run it through a deep research tool. When you've read enough, write a few lines on what you learned and send them to someone who'd care. Start tonight with whatever you've been meaning to look up. ### Sources 1. NOAA, "NOAA releases first images of nodules from waters near American Samoa," April 17, 2026. [https://www.noaa.gov/news-release/noaa-releases-first-images-of-nodules-from-waters-near-american-samoa](https://www.noaa.gov/news-release/noaa-releases-first-images-of-nodules-from-waters-near-american-samoa) 2. Dave Citron, Google, "Try Deep Research and our new experimental model in Gemini, your AI assistant," December 11, 2024. [https://blog.google/products/gemini/google-gemini-deep-research/](https://blog.google/products/gemini/google-gemini-deep-research/) 3. Henry L. Roediger III and Jeffrey D. Karpicke, "Test-Enhanced Learning: Taking Memory Tests Improves Long-Term Retention," Psychological Science 17(3), March 2006. [https://journals.sagepub.com/doi/10.1111/j.1467-9280.2006.01693.x](https://journals.sagepub.com/doi/10.1111/j.1467-9280.2006.01693.x) --- # Muscadet: the rock mattered more than we thought URL: https://srichandrasekar.com/notes/wine/muscadet-the-rock-mattered Type: Field notes / wine Published: 2026-10-25 Author: Sri Chandrasekar TL;DR: The second evening in our tasting series asked whether the rock under a vineyard changes a Muscadet. It did, more than any of us expected. Here's what we found, what we opened afterward, and what I'd pass on if you're running your own. When I wrote about [rebuilding our tastings around one question](/notes/wine/one-question-tastings), I said Muscadet was next. We poured it on October 3, and this is the write-up I promised. The evening was called [One Grape, Three Rocks](https://wine.jessri.com/tastings/muscadet/). Four wines, all Melon de Bourgogne, all from one grower, Domaine de la Pépière, southeast of Nantes. The flight came in two pairs. The first two wines share a gneiss hillside, Gras Moutons, and differ in how long they sat on the lees (the spent yeast left after fermentation): nine months for the 2024 Les Gras Moutons, fourteen for the 2022 Le Repos du Mouton. The second pair changes the rock. The 2022 Granite de Clisson grows on granite and the 2022 Gorges on gabbro, a dark, dense rock that holds water and stays cold, and both spend two years or more on the lees. I wish I could have gotten all four in the same vintage, but I was one wine short. Before the first glass, everyone made a bet: would five extra months on the lees change a wine more than a completely different rock? The room's answer was that the rock wouldn't matter much. (Melon has very little aroma of its own, which is why it's a good grape for asking the question at all.) It mattered more than anyone expected. All four wines were clearly different from each other, and I think the main difference between the rocks was minerality. That's exactly the kind of quiet note that's easy to lose in a glass, which is what [Finding the viola](/notes/wine/finding-the-viola) is meant to train. The crowd favorite was the Gorges, even though it wasn't as tight as advertised. Our own walkthrough had billed it as the tightest, most reserved wine of the four, the one built to age longest. The time pair was easy to tell apart too. I could clearly taste the difference between the Gras Moutons and the Repos du Mouton. Those two are also two vintages apart, so that difference is probably some lees time and some vintage. The rock pair, both 2022 and both long on the lees, is the cleaner comparison. ## What we opened afterward The rest of the night was less disciplined. It had actually opened with a 1996 Boizel Joyau de France, poured before the flight. After the flight came a 2021 Manincor Mason di Mason Pinot Noir from Alto Adige, which is one of the best recent Pinot Noirs I've had and the one I measure other Pinot against. Then a 2014 Daniel-Etienne Defaix Chablis 1er Cru Vaillon. My note from the night: "Great minerality and acidity on this wine. On the nose, I got lemon but a lot of lees/yeastiness. On the palate, the finish was long. I'm glad I bought several bottles!" Lees again, this time in Chardonnay. We ended with a half bottle of 2010 Clemens Busch Fahrlay Riesling Beerenauslese, and it may be the best wine I've ever had in my life. My whole note: "Laser sharp acidity balanced by amazing sweetness. Honeyed on the nose." One more bottle got opened that wasn't meant for the table. A 2022 Eleven Winery Sugarloaf Roussanne went into the risotto as cooking wine. It was better than expected, and we ended up drinking the rest over the next two nights. ![The four Muscadets from Domaine de la Pépière, with the other bottles we opened that night](/images/notes/muscadet/bottles.jpg) ## If you're running your own A few things I'd pass on from this one: - **Buy the bottles before you set the question.** Matching vintages across a whole flight from one grower is harder than it sounds. - **Get the prediction before anyone tastes.** The bet that rock wouldn't matter is what made the result interesting. Without it, "these four taste different" is just an observation. - **Taste in pairs, side by side, then go back.** That's how we ran it, and it's how both differences showed up so clearly. ![The tasting trays, and a soup I made](/images/notes/muscadet/trays.jpg) As for Muscadet itself, the sad truth is that there's a reason Melon de Bourgogne got kicked out of Burgundy. It's just not as good as Chardonnay, its sibling. I found the wines entirely drinkable and serviceable, and I thought the tasting highlighted what I wanted it to. But I'm not going to have a cellar full of Muscadet any time soon. If I want a wine to go with seafood, my answer will still probably be Chablis or Assyrtiko. ## Sources - A Cellar in Chapters, "One Grape, Three Rocks: Muscadet." https://wine.jessri.com/tastings/muscadet/ - Falstaff, Manincor 2021 Mason di Mason Pinot Nero Südtirol DOC. https://www.falstaff.com/en/wines/manincor-2021-mason-di-mason-pinot-nero-suedtirol-doc - Bowler Wine, Domaine de la Pépière, Muscadet Sèvre et Maine "Le Repos du Mouton" 2022. https://www.bowlerwine.com/wine-or-spirit/muscadet-sevre-et-maine-le-repos-du-mouton/2022 --- # My coffee setup, and how I got here URL: https://srichandrasekar.com/notes/coffee/my-coffee-setup Type: Field notes / coffee Published: 2026-09-29 Author: Sri Chandrasekar TL;DR: From an OXO grinder and an AeroPress to a Decent and an Aillio Bullet: my coffee setup, how it grew over COVID, and why roasting is what I'm chasing. When I got into coffee, I was completely happy with an OXO burr grinder and an AeroPress. Life was simpler back then. To this day, the AeroPress is still my coffee comfort zone. Then COVID happened. As a VC, I was used to three coffee meetings a day, and staying home for a year without espresso drinks just wasn't going to work. So I bought an ECM Synchronika, a dual-boiler machine with an E61 group, and a Niche Zero grinder, back when a Niche Zero was nearly impossible to get. Over time I collected more gear: a Fellow Ode Gen 2 with SSP burrs, and a steady stream of new coffee-making toys. Eventually I upgraded the two pieces that matter most. The Synchronika gave way to a Decent. I was just really impressed with how programmable it is, and for an engineer like me, it's the ultimate coffee toy. The Niche gave way to a Weber Workshops EG-1, the canonical "end-game" grinder. It's just designed so well, although it does look a bit like a microscope on my counter. Roasting started over COVID too. We were visiting my in-laws, and I found Coffee Bean Corral, a site that sells green coffee to home roasters, and a video showing how to roast small batches in a popcorn air popper. I tried it, and got pretty decent at it. That led to a Fresh Roast SR800, then a custom extension tube for it from Etsy, and then the grand-daddy of home roasting, an Aillio Bullet. ## What's on the bar today - **Espresso:** a Decent DE1PRO. - **Everything else:** the AeroPress, a Clever Dripper, a Fellow Stagg [XF] pour-over dripper, a Fellow Clara French press, and a Weber Workshops BIRD, which steeps the coffee like immersion and then pulls it through the filter with a piston. - **Grinder:** a Weber Workshops EG-1. - **Roasters:** an Aillio Bullet R1 V2, and the Fresh Roast SR800 with its extension tube. - **Kettle:** a Fellow Stagg EKG. I've largely stopped buying coffee gear. What I haven't stopped is roasting. I love light roasts, the floral, fruity kind, and nailing the temperature for them is hard. I'm still on a quest to become an excellent roaster, and I'm about to try something a little ambitious with the Bullet, which I'll write about next. If you roast light, especially on a Bullet, tips are very welcome. You can find me on [LinkedIn](https://www.linkedin.com/in/srichandrasekar) or [X](https://x.com/SriChandrasekar). --- # Roasting light is a control problem, so I'm building a controller URL: https://srichandrasekar.com/notes/coffee/roaster-controller Type: Field notes / coffee Published: 2026-09-29 Author: Sri Chandrasekar TL;DR: Light roasts punish late corrections. I'm building an adaptive controller for my Aillio Bullet that learns how it responds and gets ahead of the roast. I love light roasts, the floral, fruity kind. They're also hard to roast well, because nailing the temperature is hard, and the stakes go up with the beans. When I've bought something expensive, like a rare Panamanian Geisha or a lot with an unusual anaerobic process, I want to get everything out of it. The problem is that roasting is crude. Most traditional drum roasters have an actual flame, a gas burner, heating the drum. My roaster, an [Aillio Bullet](https://us.aillio.com/products/bullet-r1-v2), is unusual in that it heats with induction instead. But either way, you're heating the beans two ways at once: by conduction, where they touch the hot drum, and by convection, from the hot air moving through it. Outside conditions change the roast too. Humid air carries heat less efficiently than dry air, so the same settings on a different day give you a different roast. And the controls you get to respond with are coarse: on my Bullet, power, fan and drum speed, each in nine steps. Worst of all, you're always reacting. The number every roaster watches is the rate of rise, how fast the bean temperature is climbing. When it falls to zero or below late in the roast, what roasters call a crash, the coffee tends to taste flat. By the time the curve shows you a crash, it has already happened, and anything you do now is a correction, not a plan. The tools I have don't really fix this. Aillio's own software can replay a previous roast, making the same power, fan and drum changes at the same times, but its documentation is clear that it doesn't track the curve. If the room is colder or the beans are wetter, it can't adjust. Artisan, the open-source roasting software, can run a PID controller that chases a reference curve, which is a real step up from doing it by hand. But a PID loop is still correcting an error it can already see. So here's what I want to build. First, I'll do a handful of roasts to teach a model how my roaster actually responds: how bean temperature and rate of rise move when I change power, fan and drum speed, and how quickly. Then I'll give the controller targets, the curve I want to hit, and let it use that model to look ahead. If it can see a crash coming, it can adjust before the crash instead of after. And because every roast is new data, the model should keep adapting as the season, the beans and my targets change. Engineers would recognize this as a form of model predictive control. I'm going to write the code with Claude Code's help. I should be honest about the limits. I haven't built it yet. A few roasts may not be enough data to model the roaster well. And a controller can only hit the targets I give it; choosing the right curve for a particular coffee is still on me. I also couldn't find a public project doing this for the Bullet. The closest I've seen is academic work, like a [2023 University of Pretoria thesis](https://repository.up.ac.za/items/ca823592-c1fa-4cdc-8b7a-983a649055df) on predictive control for a gas-fired roaster. I'll write up what I learn as I go. If you've tried something like this, tips are very welcome. ## Sources - Aillio, "Bullet R1 V2." https://us.aillio.com/products/bullet-r1-v2 - Aillio, RoasTime documentation, "Roasts" (playback). https://docs.aillio.com/roastime/five-tabs/roasts/ - Roast Magazine, "Fire and Heat," June 25, 2024. https://www.roastmagazine.com/stories/fire-and-heat - Royal Coffee, "How External Factors Affect Coffee Roasting Conditions," November 19, 2024. https://www.royalny.com/blogs/how-external-factors-affect-coffee-roasting-conditions/ - Royal Coffee, "How to Avoid the Coffee Roasting Curve Crash," March 31, 2025. https://www.royalny.com/blogs/how-to-avoid-the-coffee-roasting-curve-crash/ - Artisan, "PID Control." https://artisan-roasterscope.blogspot.com/2016/11/pid-control.html - C. E. Bolt, "Empirical modelling in non-linear predictive control: a coffee roaster application," University of Pretoria, 2023. https://repository.up.ac.za/items/ca823592-c1fa-4cdc-8b7a-983a649055df --- # I finally bought the fun car, and promised to fix it myself URL: https://srichandrasekar.com/notes/porsche/why-i-wrench Type: Field notes / porsche Published: 2026-09-29 Author: Sri Chandrasekar TL;DR: I put off a fun car for years, convinced the maintenance would be ridiculously expensive. When I finally bought one, I promised to do the work myself. It changed how I feel about service day. My entire life, I put off buying a fun car. It was something you got once you were "successful" and "had time." Then I left Point72 after nine years and took a few months off before joining AI House. For the first time in a long time, I actually had time. I'd also spent years convincing myself that the biggest problem with a nice car is that the maintenance is ridiculously expensive. So when I finally bought one, I made myself a promise: I'd maintain it myself. That's how I ended up with three cars I work on. The first is a 2005 Toyota Land Cruiser, the 100 Series, which is my overlanding and camping vehicle. The second is a 2016 Porsche 911 Carrera 4 GTS, with a 430-horsepower, naturally aspirated 3.8-liter flat-six. 2016 was the last model year before every Carrera went turbocharged, and I'd argue the GTS is the last great naturally aspirated Porsche outside the GT3 line. The third is a 2016 Boxster Spyder, which runs a version of the same 3.8-liter engine. The promise has limits. I can't do body work, and I won't install paint protection film. But if it's electrical or mechanical, I'll take a shot at it myself. The garage has grown to match. There's a four-post lift and a sliding bridge jack, so I can get a car in the air and still get the wheels off. There's a full set of hand tools, plus pneumatic and battery-powered ones. And there's a PIWIS, Porsche's dealer diagnostic system, which lets me do the coding that jobs like a PDK service normally need a Porsche dealership for. The jobs have grown too. I started with the simple ones: rotating tires, changing oil and filters. From there I've worked up to complex ones, like the timing belt on the Land Cruiser and repacking the CV joints on the 911's front axle, which exists because it's a Carrera 4 and drives the front wheels too. The biggest change is how I feel about it. I used to dread taking a car in for service. Now I'm excited when something needs work, because it's a new skill to learn. ## Sources - Porsche Newsroom, "Porsche 911 Carrera GTS: technical specifications" (2015 model year, 991 generation). https://newsroom.porsche.com/dam/jcr:c944dc62-e433-4c19-b452-436054dcf22a/2015_911_Carrera_GTS_Technical_Specs.pdf - Porsche Newsroom, "New Porsche 911 GTS models," January 2017. https://newsroom.porsche.com/en/products/porsche-911-gts-new-models-2017-carrera-targa-13296.html - Porsche Newsroom, New York International Auto Show 2015 press kit (Boxster Spyder). https://newsroom.porsche.com/dam/jcr:042c82fb-5afe-4ffe-af0b-881c1397bbf2/PM_New-York-International-Auto-Show-2015_EN2.pdf --- # Finding the viola URL: https://srichandrasekar.com/notes/wine/finding-the-viola Type: Field notes / wine Published: 2026-09-29 Author: Sri Chandrasekar TL;DR: Most people can smell vanilla on its own but lose it in a glass of wine. I think finding one familiar aroma inside a complex wine is a trainable skill, and here's how I'm training it. I know what a viola sounds like. Put one in a room by itself and I'll recognize it. Put it in an orchestra and I lose it almost immediately. It's still there, playing its line under the violins, but what I hear is "orchestra." Smell works the same way. Most people can pick out vanilla or lemon on its own without much trouble. Put a hint of either into a glass of wine, alongside fruit and flowers and oak and everything else, and most people smell "wine." The note is in there. They just can't follow it. I think that's a trainable skill, and I've been building a way to practice it with the friends I taste with. We call it Finding the Viola, and it opens some of the evenings in our tasting series. What makes it work is splitting "I can't pick out the aroma" into smaller skills. There are five we practice. Detection is noticing that something is different at all. Classification is placing it in a family: fruit, flowers, herbs, earth. Identification is naming it more specifically. Segmentation is picking out two aromas when they're mixed together. Transfer is recognizing something related in an actual wine. Most tasting advice jumps straight to identification ("do you get the cherry?"), which is the hardest place to start and the easiest place to feel lost. So each session climbs in the same order. We start with six references: culinary aromas added to a neutral, dry Pinot Grigio, next to an unmodified glass of the same wine. With the names visible, you smell the plain wine, then a reference, then the plain wine again. Then the labels disappear. The six come back in a different order, marked with Greek letters, and everyone writes down impressions before the reveal. "Floral" counts. "Red fruit" counts. Getting the family right is progress. Next comes strength. One aroma, four glasses: an unmodified control, half strength, full strength and double strength. The first job is just to notice a difference, and only then to name it. The half-strength glass is where it gets interesting, because that's much closer to how quietly an aroma actually shows up in wine. Then mixtures. Three glasses, each with two references blended together. Usually one takes over and the other seems to vanish, and the exercise is to keep looking for both before anyone talks. Finally we reset and taste four wines, starting from our own observations with no aroma list in front of us. Only afterward do we ask whether anything familiar showed up. Two design choices matter more than they look. The first is repetition. Across four evenings there are 24 reference appearances but only 17 distinct aromas, and that's on purpose: mushroom comes back with Pinot, the Carso wines and the Jura; beeswax connects Chardonnay and Carso; green apple connects Chardonnay and the Jura. You don't learn a voice by hearing it once. The second is that the references are analogies, not ingredients. Nobody is claiming the wine contains rosemary. "I don't find it in this wine" is a perfectly good conclusion, and I'd rather hear that than watch someone talk themselves into a note because it's on the list. There's also a real limit here, and it's worth knowing. In a [1996 study](https://pubmed.ncbi.nlm.nih.gov/8934843/), both trained panelists and professional perfumers could pick out only three or four components of a complex odor mixture, and training didn't raise that number. So the goal isn't to list twelve things in a glass. It's closer to what a musician does: choose one line and follow it through the piece. That's a smaller claim than "anyone can learn to taste like a sommelier," and I think it's the one that's actually true. We're running these sessions over the coming evenings, and I'll write up what changes, including if the answer is "not much." If you want to try it with your own group, the method and a five-page attendee guide are on [our tasting site](https://wine.jessri.com/finding-the-viola/). All you need is a neutral white, a few things from the kitchen, and the patience to smell the plain wine one more time. ## Sources - A. Livermore and D. G. Laing, "Influence of training and experience on the perception of multicomponent odor mixtures," *Journal of Experimental Psychology: Human Perception and Performance*, April 1996. https://pubmed.ncbi.nlm.nih.gov/8934843/ - A Cellar in Chapters, "Finding the Viola." https://wine.jessri.com/finding-the-viola/ --- # Why I rebuilt our wine tastings around one question URL: https://srichandrasekar.com/notes/wine/one-question-tastings Type: Field notes / wine Published: 2026-09-29 Author: Sri Chandrasekar TL;DR: After a few years of tasting with friends, none of us, me included, had gotten much better at tasting. So I redesigned the evenings so each one answers a single question, with everything else held constant. I've been hosting wine tastings with friends for a few years, and I've gone through several versions of how to run them. The first version was a small group focused simply on learning about the wines of one region. We did the Rhône Valley that way, and then Piedmont. The second version was bigger, with larger groups organized around a grape or a question of value. We tasted Pinot Noir from around the world, and we ran a sparkling comparison built around whether Champagne is worth the extra money. Those were good evenings, and I'd happily do any of them again. But a few years in, I noticed something that bothered me: the overall tasting level of the group hadn't moved much. That included me. We'd tasted a lot of good wine. We weren't much better at tasting it. My read on why is that those formats changed too many things at once. A regional tasting varies producer, vintage, grape, site and winemaking in the same flight, so when two wines taste different, there's no way to say what caused it. A grape tour across continents does the same thing on a bigger map. You come away with impressions, but nothing that trains you to connect a cause to what's in the glass. And every evening was a one-off, so nothing came back around to be practiced. So I designed a series where each evening exists to answer one question. It's called A Cellar in Chapters: eighteen tastings, eight people, four wines a night, in three chapters. The first holds a single grape up to the light, the second traces a place through the glass, and the third crosses a border to see what survives the trip. The questions are the point. For Syrah, we put two Pacific Northwest wines against Côte-Rôtie and Hermitage and asked whether Syrah's personality is shaped more by place or by winemaking philosophy. For Riesling, three wines from the same producer, vineyard and 2017 vintage walk from Kabinett to Spätlese to Auslese, so sweetness is the thing that changes, and the question is when more of it improves balance and when it becomes too much. In the Jura, we'll compare Savagnin that's been topped up with Savagnin left to age under a veil of yeast. And one Pinot evening is fully blind, to see whether we end up grouping the wines by style or by where they're from. Wherever I can, the flight holds everything constant except the thing we're asking about. That's why Muscadet opens the running order: one grower, one grape, one vintage, and three different rocks underneath, so the ground is the only variable. If anyone at the table can taste terroir, that's the evening to show it. Some evenings also start with a short aroma exercise I've written about separately, [Finding the viola](/notes/wine/finding-the-viola), which practices picking out one familiar smell inside a complex wine. The two ideas fit together. The question gives each evening a reason to pay close attention, and the aroma work gives people a way to do it. We're one tasting in. Syrah came first, and Muscadet is next. It's far too early to know whether the group is getting better, and I'm trying not to decide the answer in advance. But I'm excited about it, because this is the first version built specifically to make all of us better tasters. If you run tastings with friends, the one change I'd suggest is to pick the question before you pick the bottles. Then find four wines where only one thing changes. The full series, with every flight and question, is on [our tasting site](https://wine.jessri.com/). ## Sources - A Cellar in Chapters, series overview. https://wine.jessri.com/ - A Cellar in Chapters, "Syrah North and West: Rhône vs Washington." https://wine.jessri.com/tastings/syrah-north-and-west/ - A Cellar in Chapters, "Riesling 2017: Dry to Auslese." https://wine.jessri.com/tastings/riesling-2017/ - A Cellar in Chapters, "Topped Up or Left Alone: The Jura." https://wine.jessri.com/tastings/jura/ - A Cellar in Chapters, "Pinot Without a Passport: Alto Adige vs Burgundy." https://wine.jessri.com/tastings/pinot-without-a-passport/ --- # Enterprise AI starts with the process, not the model URL: https://srichandrasekar.com/enterprise-ai-process-first Type: Explainer Published: 2026-09-29 Author: Sri Chandrasekar TL;DR: Enterprise AI works when it is applied to a process a company actually understands. Automating a broken or undocumented workflow with AI makes it fail faster, so process discovery comes first and automation second. **Process-first enterprise AI** means understanding how work actually flows through a company before applying AI to it. The AI is only as good as the process it is dropped into: point it at a clean, well-understood workflow and it compounds the gains; point it at a broken or undocumented one and it amplifies the mess. ## Why it matters Most large companies don't have an accurate picture of their own processes. The official flowchart describes how work is supposed to move; the event logs in ERP, CRM and ticketing systems show how it really moves, including the workarounds, rework loops and handoffs nobody designed. Generative AI makes it cheap to automate steps, which makes it cheap to automate the wrong steps. In a 2024 Point72 Ventures essay, Sri and Ryan Fauber argued exactly this: AI won't fix enterprise process problems on its own, and companies need process intelligence before they automate. ## How it differs from nearby ideas - **Robotic process automation (RPA)** scripts existing steps as they are. It is automation without understanding, so it inherits every flaw in the process. - **Copilots** assist an individual inside one tool. They make a person faster at a task but don't question whether the task should exist. - **Process mining** reconstructs the real process from system event logs. It is the discovery layer, not the automation layer. - **"Agents"** are a delivery mechanism for automation. Sri listed "agentic" as the most overhyped word of 2025 in GeekWire's investor predictions, while calling process mining one of the categories AI brings back to life. ## What to evaluate 1. **Is the process observable?** Is there an event trail (timestamps, case IDs, actors) or does the work happen in email and spreadsheets? 2. **Is the process stable?** Automating a process that changes monthly locks in a moving target. 3. **Where is the real cost?** Rework, waiting time and exception handling usually cost more than the "happy path" steps that are easiest to automate. 4. **Who owns the outcome?** Someone has to be accountable for the process after the AI is in it. 5. **Does domain expertise sit with the product team?** Sri has pointed to product managers with deep domain knowledge as the key hire for enterprise AI. ## A practical method 1. **Discover.** Mine event logs, and interview the people doing the work, to map the process as it runs today. 2. **Fix.** Remove the rework loops and unnecessary handoffs first. Some of the gain comes before any AI. 3. **Instrument.** Define the metrics (cycle time, error rate, cost per case) that will show whether automation helped. 4. **Automate selectively.** Apply AI where the process is now stable and the value is measurable, and keep people in the loop at the judgment points. ## Where it goes further: AI that prompts people Once a process is understood, AI can do more than execute steps. In "Clunky copilot or benevolent boss?", Sri argued that workplace AI should behave like a chief of staff: it knows the priorities, tracks the work and prompts people about what needs attention, rather than waiting for a person to prompt it. ## Common misconceptions - **"The model is the hard part."** For most enterprise deployments the hard part is data access, process clarity and change management. - **"Automation always saves money."** Automating an inefficient process can raise costs by making the inefficiency faster and harder to see. - **"Process mining is old news."** The technique predates generative AI, but AI makes both the discovery and the follow-on automation far cheaper. ## Limits Process mining needs good event data, and much knowledge work leaves no clean trail. Some valuable processes are too variable to standardize. In those cases AI is better used to assist people than to automate the flow. ## How Sri's work relates Sri has written about this theme in two Point72 Ventures essays (2024) and returned to it in public predictions (2025). Several of Sri's investments apply AI to a specific, well-defined workflow, for example healthcare revenue cycles (Adonis) and in-house legal work (Luminance). See [Writing & talks](/writing) and [About](/about). --- # Voice AI in customer service: when callers prefer the machine URL: https://srichandrasekar.com/voice-ai-customer-service Type: Explainer Published: 2026-09-29 Author: Sri Chandrasekar TL;DR: Voice AI for customer service uses speech recognition, language models and speech synthesis to handle phone calls end to end. It works when callers would choose it over waiting for a human: fast, accurate, and able to finish the job. **Voice AI for customer service** is software that answers and resolves phone calls in natural conversation. It combines speech recognition, a language model that understands the request and decides what to do, integrations that let it take action in the company's systems, and speech synthesis to respond. The bar it has to clear isn't "good enough to deflect calls." It is good enough that callers would rather talk to it than wait for a person. ## Why it matters The phone is still where customers go when something is urgent or complicated, and call centers are expensive, hard to staff and hard to scale up and down. Legacy phone trees (press 1 for billing) solved the cost problem by making the experience worse. Voice AI is worth deploying only if it solves both at once. That was the thesis Sri and Daniel Kwak laid out when Point72 Ventures backed PolyAI in 2019: build voice agents good enough that callers prefer them. ## How it differs from nearby ideas - **IVR phone trees** route callers through fixed menus. They sort calls; they rarely resolve them. - **Chatbots** handle text. Voice adds hard real-time constraints: latency, interruptions, accents and noisy lines. - **Agent-assist tools** help a human agent during a call. Voice AI handles the call itself and hands off when it should. - **Human-in-the-loop systems**, like the machine-plus-human translation model behind Unbabel (another Point72 investment Sri led), blend automation with people for quality. Good voice deployments borrow the same idea through clean escalation to humans. ## What to evaluate 1. **Resolution, not deflection.** What share of calls is fully resolved without a human, and do those customers come back? 2. **Latency and turn-taking.** Pauses and talking over the caller destroy trust quickly. 3. **Integration depth.** Can the agent actually do things (rebook, refund, update an account), or can it only talk? 4. **Escalation quality.** When it hands off, does the human get the full context? 5. **Caller preference.** The real test is whether callers choose it, measured in satisfaction and repeat calls. ## What changed with large language models Earlier voice systems depended on hand-built intents and dialogue flows, which were brittle outside the paths designers expected. Large language models handle open-ended requests far better, which widens what can be automated. They also add new risks: confident wrong answers and actions taken on a misunderstanding. The fundamentals (fast turn-taking, deep integrations, clean escalation) matter as much as before. ## Limits Calls involving emotional distress, complex judgment or regulatory risk still need people. Voice AI also inherits the quality of the systems and policies behind it: an agent can't resolve what the company's own processes can't. That links back to [enterprise AI starts with the process](/enterprise-ai-process-first). ## How Sri's work relates Sri co-wrote Point72 Ventures' PolyAI investment thesis in 2019 and served on PolyAI's board from 2019 to 2025 (UK Companies House). Sri also led Point72's investment in Unbabel and spoke about automation in the language industry at SlatorCon in 2019. See [Writing & talks](/writing) and [About](/about).