Work

Mission Control turns what expert workers know and never wrote down into AI agents that run inside the customer's own environment. The customers sit in defense, aerospace, energy, manufacturing and similar industries, so the constraints are hard: Microsoft identity, mTLS, customer-data privacy, licensing. I joined in April as one of two engineers. The job turned out to be about half engineering and half running the customer engagements the engineering was for.

It is the same bet as "Taste" is Not Ineffable: what an expert knows can be written down, and once it is written down, a machine can apply it.

Customer-facing workflows

Most of the year went into the workflows customers actually touch: an AI assistant reachable from a normal Microsoft Teams chat with no separate account to set up, research briefs where every claim is checked against its source, and the human-approval boundary that keeps any of it from sending or changing anything on its own. The specifics stay with the customers.

The certificate that would not work

A high-security customer had a browser-certificate failure with no clean reproduction. I turned it into a reproducible mTLS lab, traced the root cause across worker processes, and drove one-shot secret handoff plus concurrent identity isolation. I proved the underlying Chromium path locally and kept the customer's final acceptance explicit rather than assumed.

Licensing and reliability

Self-hosted customers need a license system that fails safely, so I built the signed licensing and enforcement foundation, with enforcement that degrades in phases instead of stopping cold. Underneath it, a reliability program: an audit of the platform, fixes in batches, a deletion pass, and the QA harness the team used through the summer, with recorded fakes for outside systems, a nightly run, and bugs filed automatically with their evidence.

engOS

In May I designed engOS, an AI-native engineering control plane: human intent and authority bound to implementation candidates, audit findings, evidence and release state. Seven weeks in, measured product output had gone from 36 file-touches a week to 128, and rework from 12.4% to about 5%. By September the mandatory version cost more in review time than it returned, so I replaced it with a short shared engineering loop, kept the proof requirements, and moved the tool out of the product repository.

Across all of it, I direct AI agents through implementation and adversarial review, and I own requirements, architecture, verification, and acceptance. Customer names, product details and the numbers behind them stay with Mission Control. Ask me in person.

Most professional relationships go quiet by accident. FollowUpEasy connects to Gmail, reads the whole mailbox, judges every relationship, and maps the ones worth keeping: where each one stands, what is owed, whose window is open. Then it drafts the follow-up in your voice. You approve every message. It never sends on its own.

I built and operate all of it: discovery, product design, architecture, Gmail OAuth, encrypted credentials, PostgreSQL, model inference, grounding, background workers, billing, analytics, deployment, debugging, human approval, and production operations.

I started prototyping in January 2025 and launched in 2026. The early users have been direct. One returning tester found the relationship map more useful than the drafts; another asked to add a colleague after a trial. The less visible work is diagnosing failed scans and getting onboarding to work on the first try.

The hard part is not the model. It is deciding what the product is allowed to do. It has explicit boundaries around unsupported output, sensitive relationships, duplicate work, spend, and sending authority, so it stays useful without silently taking a consequential action. One rule I keep coming back to: when the system cannot judge a relationship, it must fail open to "we don't know," never closed to "not a real relationship." The second kind of failure deletes real people from the map.

The thinking behind it is in While You Were Working, Half Your Network Disappeared.

Claude Code's default context management was broken for long projects, so I built a persistent memory system for it. The design is a write-gated, tiered architecture with correction propagation: a fact has to earn its place, a correction propagates everywhere the old claim lived, and stale or unauditable context gets limited instead of accumulating.

I built the first version with Claude Code in three hours and launched it on Hacker News. Then I managed the product cycle in public: triaged community feedback, rewrote the docs based on user complaints, answered architecture questions, and opened design issues for the most requested features. 200+ GitHub stars.

I built Cookie Calm to take a small, repetitive annoyance out of browsing. The Chrome extension automatically rejects optional cookies and dismisses supported promotional popups, including newsletter signups, discount offers, and chat greetings.

I built on Consent-O-Matic's open-source rules and added guarded actions, popup dismissal, and per-site controls. Unknown consent forms stay visible; accepting cookies is an explicit opt-in fallback. Everything runs locally, with no account, telemetry, or remote AI. Open source under the MIT license.

The argument about whether AI helps people is mostly conducted with anecdotes. Because of AI is a registry of the cases that survive four tests: the benefit reached real people, it was measured against something, the comparison shows the AI made the difference, and we know when it happened. An entry that fails any one of them does not go in. It is built for readers who have been burned by AI hype, so its credibility comes from what it excludes.

Every entry carries three separate grades, for evidence, attribution, and timing, and keeps the date the benefit happened apart from the date the paper came out, which is where most “AI breakthrough” headlines go wrong. A “Not because of AI” section records the clean nulls, because a trial that found nothing is worth as much as one that did. An automated run reads the trial registries and journals every day and drafts entries from primary sources. Models from more than one vendor check each draft against the source it cites, and the checks and disagreements are published so a reader can inspect them. No human reviews entries. Open source: MIT code, CC BY 4.0 data.

The essay argues that AI slop is not a collection of bad habits but a single failure mode: uniformity disguised as craft. A technique spread evenly across a piece is the tell, whatever the technique. The tells come in three layers, from false binaries and unpriced claims, through metronomic rhythm and three-item closers, to prose that never touches a constraint.

The tool is the essay made executable. It generates, grounds every factual claim against a source (one unsupported claim fails the text), runs a regex precheck and a rubric score, and rewrites with the feedback. Rubrics are YAML. Built-ins cover email, essay, blog, grant and newsletter. It's named for Ramsay Brown, the friend whose critique of a draft started it: he identified the problem and I built the framework.

Employee #180. Ten years across product operations, from startup chaos through IPO to 2B daily queries.

I spent the bulk of my Facebook career on Search, including the Graph Search era, what Mark called Facebook's "Third Pillar," announced eight months after a rough IPO as proof the company could build beyond News Feed. It was one of the most technically ambitious projects the young company had undertaken: a privacy-aware index, NLP engine, and personalized search across 240 billion photos and 1 trillion connections, built by a team of 50+ led by the creator of Google Maps.

Underneath the launches I built and ran the readiness system: test plans, cross-office coordination, outsourced QA, and bug triage across 100+ engineers. Launches went out across Product, Engineering, Design, Legal, Privacy, Safety, PR, and Operations. That was the infrastructure that kept releases from blowing up.

The product had a fundamental usability problem the team was too close to see. The design concept was a "magical page title": no white search box, no search icon, just a clean blue line across the top of every page. It was an intentional, ambitious design thesis. But users couldn't find the search bar. We were flooded with bug reports: "my search bar is gone." I kept pushing eng, design, and product: just make it white. There was real resistance because the blue bar was the concept. We tested it. White outperformed the next best option by more than 2x.

That experience taught me something I've carried ever since: when the people building the thing can't see what's obvious to the people using it, someone has to be willing to say the simple, boring, correct thing out loud.

After a decade inside a company that had everything, I went and built something with none of it: an audio-first social product, with a technical cofounder, Dennis Stücken, as CTO and Eleken as design partner. I owned the product, customer discovery, hiring, fundraising and the calls. We shipped the iOS app in November 2019, and it won the Anything Else category at Product Hunt's 2019 Makers Festival with Snapchat. It didn't work as a company, and the experience changed how I operate. I learned what it feels like to have no leverage and no safety net. It gave me a sharper lens on what founders need versus what they say they need, which defined how I invested for the next four years.

I started a deep tech and AI seed fund 18 months before ChatGPT launched, when most LPs weren't interested in the thesis: back founders who manipulate matter, bend energy, or rewrite biology. Early-stage companies across AI infrastructure, semiconductors, robotics, biotech, space and energy. Companies in the portfolio went on to raise from Sam Altman, Alex Wang and Lightspeed.

The part I'd do again was the work with founders through high-stakes decisions: operational focus, capitalization structure, when to hold and when to move. The kind of calls where five investors give five different answers and only one is right.

One example. Shapes: a 2021 investment in AI companions, over a year before ChatGPT. Cold inbound from the founders, stayed in contact, invested when AI friends sounded like a toy. Lightspeed led the seed round.

Two recent ones. A founder's fundraising memo claimed more than the underlying agreements supported. I went through it against the contracts, separating nonbinding arrangements from signed ones and conditional rights from downside protection, and the corrected memo is the one that went out. Another founder was negotiating a first enterprise project; I argued for a bounded scope, payment milestones, and keeping the company's own IP out of it.

We stopped raising and making new investments in late 2025. I still support the portfolio and its investors.