Own product · Agentic build
Shipped 32,000 lines without hand-writing one of them.
A self-hosted AI dungeon master, built end to end by Claude Code. It holds together because the model is never the authority — twice over. In the product, the app owns every fact and the LLM only proposes changes a validator can reject; a second pass then audits the narration for changes it described but never declared. In the codebase, the same rule: 1,567 tests, a pre-push gate and per-issue worktrees decide what lands, not the agent. The rails are the work — the agent is just fast.
grimhold.world ↗ Series A · Fintech
Cut their LLM bill by 71% without losing accuracy.
A growth-stage fintech was bleeding $18k/month on a single GPT-4 endpoint. We rebuilt the eval harness, routed 80% of traffic to smaller models, and kept the top-line accuracy within 1.5 points.
Small SaaS · Product validation
Killed the features customers wouldn't use — before a line of code.
A small SaaS was shipping on gut feel: fat roadmaps, lukewarm uptake. I set up an agentic validation loop that stress-tests every feature idea from multiple angles — user, market, build cost — before anyone opens an editor. They now ship thinner slices that customers actually adopt.
10-person team · Health
Built an agent that closed 4 hires worth of work.
A clinical-ops team was drowning in intake forms. We deployed a focused agent with strict evals and a human-in-the-loop checkpoint. It now handles 70% of intake autonomously.
B2B CRM · Eng automation
Turned Linear tickets into reviewed, deployed PRs — hands on the wheel.
A CRM team wanted AI to clear their backlog without it going rogue. I built a 13-phase pipeline on Claude Code: parallel risk analysis, three independent AI reviewers cross-checking every change, and six human approval gates so nothing merges or ships without a sign-off.