I build AI systems for the people who run the work, and take them to production.

I joined to sell, and ended up building the system the company runs on: a hub with its own agent over about 190 files of pipeline, account plans, battlecards and playbook, with 24 scheduled agents (inbox drafter, follow-through, investor desk, PR tracker) and hourly email and Slack checks.

The demo engine is what prospects notice. Give it a brief and it builds a fictional company's own system with an assistant inside that acts as whoever is signed in, on top of 13 base demos I built for HR, MSPs, restaurants, mortgage, payroll and leave. A tester agent runs up to 14 scripted passes against every build before a person sees it.

Log a meal from a photo or a sentence. It tracks every nutrient, tells you which ones you keep missing, adjusts the program, and reads the scale and your pantry from pictures. SMS coaching, an admin dashboard for support, and its own Claude connector so a coach can read and write the app's data without touching the database.

Click any depth and you get the pressure, the temperature, what lives there, what happened there, and the projects you can join or fund. Under the descent is a platform I built in 24 days because I couldn't stop: live marine alerts, a live globe, the trenches, an ocean history timeline, fisheries, a diver's path, surf, community uploads, careers and volunteering.

Feed it CSV exports from the CRM, the sequencer, the calendar and the call recorder. It finds follow-ups that never happened, deals with one contact and intent nobody answered. The report is a chat with an analyst: tell it your real win rate and the whole model re-prices with your number labelled as yours.

Chat is the interface: "find me fifteen buyers at this account" or "draft the follow-up from today's call" happens inline, and a background worker has the day's plan ready before anyone sits down. I ran my own outbound on it every day for three months.

What broke, since that's the useful part: the web service was pinned to a frozen image tag, so for one session every web change silently never shipped while the worker did. The rule that came out of it, verify the running image digest after every promote, is in every pipeline I've built since. I retired Anchor because I came to think companies solve this in-house or buy it, and the later work follows from that.

It interviews you by voice to learn how you actually talk, sets content pillars, and measures whether a post did anything instead of counting likes. The data model is five primitives (a voice, whoever operates it, whoever pays, a campaign, the intelligence that comes back), so a solo founder, an exec with an EA and an agency running twenty clients are configurations, not code branches.

Sign-ups didn't come, so I paused it with a one-line switch that archives every job and deletes nothing. It worked; people didn't want it enough.

A multimodal model works out what you're doing, then a build route and a buy route run in parallel into a decision engine that returns vendors, a price, total cost of ownership and savings, with sliders you can drag. Capture stays in the browser; the server only sees sampled frames.