Case studies
Projects I built end to end: the problem, the architecture, and the tradeoffs.
LLM code reviewers comment on lines that aren't in the diff and pad reviews with low-confidence nitpicks. Nash AI is built around those two failures.
A GitHub App that reviews pull requests with inline comments, per-repo config, a Redis-backed worker queue, kill switches, and an eval harness.
News feeds optimize for alarm. Gentle Stream is a feed of real, recent, uplifting stories, ranked per reader without an LLM call on every scroll.
A news product that surfaces only uplifting stories in a broadsheet UI, fed by a three-agent pipeline with an LLM-free ranker on the request path.
Long-running agents are hard to audit after the fact. This runtime records every decision as a tamper-evident event and limits actions up front.
A ledger-first agent runtime: every decision is a hash-chained, verifiable event, and the actions an agent can take are constrained before sampling.
