John Hendricks LinkedInGitHub

I turn language models into systems that do real work.

I'm John, an applied AI engineer. I build agents, pipelines and automation that take over slow, repetitive work, and I test them until they can be trusted. Much of it began as tools for research and journalism. Every drawing on this page shows something I built.

A satellite constellation, mapped
The kind of research my first tools were built for. I pulled public tracking data for 168 satellites and found they all fly in one shell about 1,145 km up, in two rings. Each dot is a real satellite, with its orbit height drawn to scale. Drag to turn it.

Selected work

One of each kind of system I build most: an agent product, a research assistant and a pipeline. Tap a drawing to run it.

1Agent Rooms

Rooms where builders chat together, each with a private AI agent at their side. Promoting an agent's output opens a real pull request, other people fork and remix it, and the room adopts a winner by merging at a pinned commit, so credit follows the code.

In the drawing, the lowest line is the repo. A proposal branches off it, a remix branches off the proposal, and the red merge is the room adopting it. Before any person used it, five agent personas ran that whole loop unattended: 24 pull requests opened and 12 merged, for about three cents.

Next.js, Postgres, GitHub App, OpenRouter. Deployed. Code on GitHub

2Research Assistant

Ask it a question and it answers from 34,000 documents, with a numbered source on every claim. Seven kinds of sources flow in from pipelines I built, and it indexes everything again every four days with nobody touching it.

The hard part is finding the right 8 passages out of 149,000. It rewrites each question up to four ways, searches each version by meaning and by keyword, and merges the results by rank, so a passage several searches agree on rises to the top. When the documents fall short, it says what is missing.

Python, Claude API, Chroma, SQLite, FastAPI, and a web interface written by hand. The data is private, so the public code ships with a sample library and the test that scores it. Code on GitHub

3LLM Source Monitor

Keeping up with one subject across RSS, X and Telegram means hours of scrolling. This collects the last 48 hours on a schedule, groups the posts by topic, writes a digest for each topic and a summary across them, and keeps a wiki of the people and organizations it comes across.

It is built to run with nobody watching. Every model call streams and retries, and if the login for one platform fails, the run skips that platform and still finishes.

Python, Claude API, feedparser, Telethon. Code on GitHub

More builds

Four smaller projects, each built around one specific problem.

4Research evidence engine Works out when three names mean one company, and remembers where every claim came from. Python, SQLite. Private repo.
5Lead Qualification Pipeline Sorts a contact form inbox into hot, warm, cold and junk, then drafts the reply. Code on GitHub
6NYC Transit Delay Predictor Predicts subway delays, and tells you when it is only guessing. Python, FastAPI, Expo.
7Shiori Writes a brief for a Japan trip, modeled on the booklets travelers make before they go. Next.js, Claude.

Published analysis

8Ships that change identity

I followed 12 ships through public identity records and found that 7 had come back under new names and flags. Each row is one ship over time. Blue is its original name, and red is the same ship after it came back under a new name and flag.

Global Fishing Watch API. Published analysis, September 2026. The counts are real. The timing in the drawing is illustrative.