Builds

I built it for Mangrove’s Game Night hackathon (September 25 to 27, 2026) in about 11 hours with Claude Code. Claude wrote the code and drafted most of the text; I directed the design, playtested, and made the calls.

You read the work, spend a few scarce control tools (set to their real-world accuracy) where they count, and decide what ships. The week escalates through sycophancy and false claims, prompt injection, reward hacking, data exfiltration, collusion between agents, sandbagging, and a model that behaves differently when it thinks no one is watching. Each threat is tied to a real paper or incident, cited in the game.

It’s also a roguelike: a hidden adversary personality, events, clients, and upgrades change from run to run, and there are 3 difficulties, 4 labs, 5 model versions, and a daily shared run. For learning, it has tours of each screen, hover definitions for every technical term, a field guide to research approaches and their tradeoffs, and a short field test after a run. It’s Vite and TypeScript with no framework, and every visual is SVG drawn in code.

I’ve used it to direct AI-assisted development of more than 10 applications, in fitness, research, education, accounting, legal and compliance, cybersecurity, and marketing.

OPTIMADE gives many materials databases the same query language, so one search can cover several of them. The hard part is everything after the search: keeping track of which database each row came from, versioning a dataset so a paper can cite it, and getting it into pandas or Jupyter. When a database is slow or down, Alloybase shows the results it has and says which sources are missing.

I took over my company’s chatbot when our full-stack developer left and started over in Langflow and custom Python. We’ve since shipped tailored versions for two product launches.

The main thing I learned is where the model should and should not sit. My first version was an all-in-one agent that made decisions and ran the retrieval flows itself, and it burned through tokens and gave a mediocre experience. Now a structured-output model identifies the intent, code runs the retrieval that intent needs, and every AI feature has a defined failure mode instead of a guess.

Prompts are versioned and tested against a fixed bank of common member questions, and I think that maintenance process matters as much as the first build.

Coaching runs through a multi-agent language-model pipeline. Prompts ship as versioned prompt packs, and A/B tests between versions are stored in the database, so two versions of a prompt can be compared directly. It has subscription billing and runs on DigitalOcean.

Both are internal tools I built at my day job, so there is nothing public to link. Launch setup went through three iterations: a workflow redesign around project phases, a helper script, then a full Google Apps Script orchestrator.