Building
Most of what I make is Groupthink. I started the company in 2021, shipped the first version of it in February 2023, and since May 2026 I have built the whole thing by myself, with AI doing the work a team of engineers used to do. Here is the story, with the numbers.
The numbers
October has 11 merged so far. It is too early in the month to say where that lands.
Engineering merged pull requests
Operations commits
Engineering is the API, the web and desktop apps, the iOS app, and the marketing site. The dip in the middle is what it cost to lose a team, and the recovery on the right took four months of learning how to direct agents well.
Operations is the repositories that run the company rather than build the product: product holds the roadmap and the specs, marketing holds the campaigns, and company holds the strategy, the decision log, and the runbooks. It is counted in commits, because those repositories take direct commits and a pull request in them would gate nothing. Counting their pull requests gave fifteen straight months of zero while the writing was in fact going on.
Different units and different scales, so read each chart against itself rather than against the other. The repositories do not overlap.
counts refreshed daily from github · last updated 1 day ago
If more pull requests only meant thinner ones, the size of each would have fallen as the count rose. It did the opposite.
| Measure | A team of five November 2024 to July 2025 | Solo, with agents May to September 2026 |
|---|---|---|
| Merged pull requests | 1,614 | 2,891 |
| Lines changed | 269,383 | 1,024,039 |
| Lines per pull request | 167 | 354 |
| Commits per pull request | 1.8 | 1.8 |
| Median commit | 24 lines | 87 lines |
Nearly four times the code changed, in five months instead of nine, on pull requests more than twice as large, with no more commits inside each one. The median commit is the number that settles it, because a median cannot be moved by one big import.
api and frontend only, which is 3,197 of the 3,762 pull requests merged in 2026 through September 30, so the pull request counts here run below the chart. This measurement stops there. Lockfiles, build output, vendored code and generated files are excluded, and so are the three largest commits of every month, so that one bulk import cannot carry a month. measured October 1, 2026
1,261
merged PRs
September 2026, the best month in the company's history
Nearly double August, at 707. June before that was 604. With five people the record was 287, in February 2025.
$22
per merged PR
All-in company cost in September
It was $430 per PR in June 2025, running a team of five on $104K a month.
$200
per month
What the engineering labor costs
A flat Claude plan. The same usage at list API prices would have run about $5,200 in June.
3,762
merged PRs
In 2026, January through September
Across the API, the web and desktop apps, the iOS app, and the marketing site.
Every figure here is a merged pull request count from GitHub or a line from the company's own books. Nothing is estimated except the pace, which is labelled where it appears.
What reached production each week since I started building solo. It only lists what I have written about or checked in production.
Week of September 7
When you log an interaction you can say how it happened, and it can span several days. The contacts grid now shows last meeting and last contact separately.
Week of August 31
In the desktop app, hold the fn key and speak. Your words are typed wherever your cursor is.
One field at the top of the sidebar takes whatever you want to log. Everything you capture has its own Captures page in the left navigation.
Week of August 24
You can share a link that shows your open times across all your connected calendars. People pick a slot on a public page, with no account needed.
Groupthink can copy busy blocks between your connected calendars. The copies hide the meeting details.
Week of August 10
Action items Groupthink is unsure about now appear as suggestions you accept or dismiss. You can also pick an owner for unassigned items.
Week of July 20
You can set a check-in rhythm from a contact's page.
Week of July 13
Ask a question about a contact and the answer now shows numbered sources for each claim.
Week of July 6
You can connect more than one Outlook account. Each calendar also has its own auto-join switch for the notetaker.
The Today page lists people you are overdue to talk to, based on the cadence you set for each.
Week of June 29
You can connect another Google account in Settings. Its calendars are grouped under that email, and events on both accounts show once.
Week of June 15
Groupthink now has projects. Groupthink suggests new ones that you accept or dismiss.
A relationship page has a question box. Ask something about the person and get an answer from what Groupthink knows.
You can send a meeting action item to GitHub or Linear as an issue.
Week of June 8
A relationship page now opens with a written summary of where things stand. One timeline holds its meetings, notes, and new facts.
Press Command K on any page and type one line about someone. Groupthink files it to that person.
The Capture page now takes images, PDFs, and audio files. Groupthink reads them and you file what it finds to a person in one tap.
Action items from a meeting now show in its recap. A new view lists your open action items from every meeting.
Week of June 1
Each person's page now shows past meetings, facts you can add and edit, and a list of open topics to raise next time you talk.
You get a personal address for forwarding meeting notes from any notetaker into Groupthink. It lives in Settings, under Ingestion.
You can choose several calendars, and your meetings list pulls events from all of them.
Week of May 18
Through the Claude connection, Claude can now write a note onto a person in Groupthink and look up past meetings.
Meeting notes and recaps show a row of the people linked to that meeting. You choose which people a meeting's notes go to.
Interactions you log by hand, like a dinner that was not on your calendar, now count toward when you last saw someone.
Week of May 11
Claude and other AI assistants can now connect to Groupthink. They can read your meetings and relationships and send a notetaker to a meeting.
A new API Tokens tab lets you create the tokens that connect Claude Desktop, Claude Code, or Cursor to your account.
Week of May 4
Clicking someone on the Relationships list now opens their own page. It shows when you last met and whether you are keeping up with them.
What Groupthink is
A relationship intelligence platform. It makes the dynamic between you and another person legible enough to act on, so your next meeting with someone is better than your last.
Not the meeting and not the task. A relationship is a system that exists in the space between two people, and inside it the load-bearing points are the commitments you make out loud. Those are the checks we write with our mouths that our workload cannot cash.
This software has always been sold to companies, to manage their people. Groupthink inverts that. You are a node across many overlapping and temporary organizations, and the record of your side of them should survive changing jobs, tools, and AI vendors.
Every meeting, every paste, every detected name deepens what the system knows, whether or not you remember to file anything. A notetaker's summary is identical on meeting one and meeting fifty. This one gets better every day you do nothing.
What it is not: a meeting notetaker. Meet, Teams, and Zoom all transcribe and summarize for free. That was the whole problem, and it is where this story starts.
Five years, one category collapse, one team reduction, and a rebuild that worked. The unflattering parts are in here too, because without them the rest does not mean anything.
October 2015
Six years before the company exists, a post asks for a personal CRM for grown-up friendships: "My Mom had this pretty blue book where she'd record all of this stuff about her friends and family. I want something like that, but for the modern age." It ends: who will build it?
Read the 2015 post →2021
The company is incorporated as First Party Labs. Groupthink's first line of code is still two years away.
February 2023
First commits land in the API and web repos on February 2nd. A tool for running better meetings: agendas, notes, the things a team needs to not waste an hour.
2024
Google Meet, Teams, and Zoom all ship native transcription and summaries, free, with no bot to add. The most common objection from new signups becomes the one we could not answer: why would I use this instead of what is already in my meeting tool?
February 2025
287 merged pull requests in a month, from five people. It stood as the record for fifteen months.
August 2025
Burn goes from $104K a month to about $70K, then $38K. Output goes with it. This is the part of the story that is not a growth chart.
December 2025
17 merged PRs that month, then 15 in January. Two people, one of them part-time and reviewing only. The cost per shipped PR was worse than it had ever been with a full team.
February 2026
Claude Code becomes the engineering team. 98 PRs in February, 113 in March, 127 in April. Slower than the team era, and improving every month.
March 2026
Groupthink stops being a meeting notetaker and becomes a relationship intelligence platform. Meetings are inputs. What you know about a person after twenty of them is the output.
May 2026
The last reviewer comes off. 424 merged PRs that month, past the old team record, at $71 per PR against $430 eleven months earlier.
June 2026
More than double the best month the team ever had. The binding constraint stops being engineering throughput and becomes deciding what is worth building.
Now
The product surface is an advisor that reads your record and tells you where a relationship actually is, and what the next move toward what you want from it looks like.
People assume the interesting part is the model. It is not. The interesting part is the scaffolding around it, which took about six months to get right and is still changing every week.
Strategy, decision log, runbooks, meeting notes, the operating principles, the relationship record. Not because it is tidy, but because it is what an agent can read. A page of English is a job description.
A milestone exists in all five repos at once: product, API, web, website, marketing. It is the single answer to what ships, when, and what goes out with it.
Several at a time, each in its own git worktree so they cannot collide, each claiming its issue first so two never build the same thing. That collision cost a duplicated feature once, which is why the claim step exists.
Every PR gets reviewed before it merges. A human reviewer was the last role to go, in May 2026. The review is the final gate, not a pre-flight check.
Merged is not done. Deployed is not done. Done is the post-deployment checklist actually run against production, with the results written down. CI cannot exercise production, so a checklist that is never executed is theater.
The hand-built version will always be ahead of what ships. That is not a gap to close, it is the roadmap arriving early. What works here gets built for people who will never open a terminal.
Checked September 2, 2026
Connecting a calendar has to leave you with a record that is already populated. Today it can leave you on an empty screen.
Catching what was actually promised out loud, in the room, and holding it somewhere you will see it.
Your real CRM is your calendar and your texts. Email landed first. This is the next signal.
An MCP surface your other agents can read the record through. The point of owning it is being able to hand it back to yourself.
Capture that works without a bot joining the call, on a machine you control.
The one screen that has to be worth opening every morning.
The honest part
A throughput chart is not product-market fit. I can ship more in a month than a funded team of five could, and I still spend most of my thinking time on whether the thing I am shipping is wanted. In July I wrote this down while sitting on a balcony in Rome, and it is still where I am:
Paying customers, low churn, low growth, low burn, and I am having a lot of fun figuring out how to build and launch major features completely solo with the help of AI. The only problem is, it is lonely and I am not sure I am making something people want.
The constraint stopped being engineering a while ago. It is deciding what deserves to exist, which no agent has done for me yet.
Nineteen years of building, in order. The links go to what I wrote at the time, and the list ends where the top of this page begins.
2007–2009
My first startup job, on the data team of the Seattle company building Whrrl, collecting and cleaning the place data behind a location-based social app. On the side I was the events editor at Seattle 2.0, covering the startup scene one event at a time. I left with no plan beyond working for myself.
Pelago, day two (2007) → Goodbye Whrrl, Hello World (2009) →
2009–2012
Hired to do support, blog posts, and video when the company was still working out of borrowed office space; I asked for a senior title on the business card so I could get meetings. Over three years I built the marketing team from zero to eighteen people, launched Twilio SMS, and made the Forbes 30 Under 30 list along the way.
We launched Twilio SMS (2010) → I don't do that job anymore (2013) →
2012–2013
Y Combinator, Summer 2012. Raised $1M, grew to ten people, and learned the idea was wrong. We killed it rather than run a zombie, and the post about that decision outlived the company.
Starting Referly took me three years (2012) → Zombie Startups (2013) →
2013–2017
The deal intelligence company, started with the $350K left in Referly's bank account. Union Square Ventures was the first paying customer, the Mattermark Daily peaked at 100,000 readers, revenue reached about $3M ARR, and the team peaked at seventy. The sale to FullContact in December 2017 did not clear preferences, so common stock, mine included, was wiped out. I stayed six months at the acquirer to land the plane.
Introducing Mattermark (2014) → This Storm Is You (2018) → What the sale actually paid (2019) →
2018
Self-funded, after nine months of M&A. Over a hundred books, angel investing, and a partnership at XFactor Ventures backing female founders.
2019–2021
Ran Meltano as an internal startup, hired my replacement, and watched it spin out with a $4.2M seed led by GV. Then built the growth marketing team, reporting to the CEO for my first year. I left shortly before the IPO.
Can I find my next great investment? (2021) → What is a week end? (2020) →
2020
One public company analyzed each day, alphabetically through the Vanguard Total Stock Market Index, as a paid Substack alongside the GitLab job. The run now lives in this site's gated archive.
2021–2023
Developer-first marketing analytics, built with Jonathan Kressaty. Incorporated in August 2021, funded within 48 hours of tweeting about it, and backed with a $5.7M seed in 2022. We never found product-market fit, so we pivoted. The pivot is the company this page is about.
A Chip and a Chair (2021) → A Robot Wrote This Subject Line (2022) →
How the work actually gets done — the systems, the rules I've written down, and what breaks.
One-on-one with founders, executives, and operators, on building companies and on the career questions underneath that.
How I coach →Twenty years of it, from the Mattermark era through now. All of it lives here rather than on a platform that can change its mind.
Read the archive →A whole-food, plant-based kitchen run to restaurant standards, at 5,280 feet, with the same systems thinking pointed at dinner.
See the kitchen →