Selected work
Each project carries the same through-line: state the capability, then state the leash. They span production work in a high-stakes environment, open-source repos that are public and verifiable, and agent teams that run real businesses against real money.
Production
2 case studiesProduction agent systems in a high-stakes environment
I design and run production agentic systems with guardrails: multi-agent work where capability is bounded by hard safety limits, with a human on every consequential decision.
Real stakes, bounded by designLocal-first practice assistant
A Windows desktop app for a solo therapist, built on the Agent SDK, that handles session notes, scheduling, intake, and billing. The sensitive data is encrypted and never leaves the device. Open-source and anonymized.
Privacy is the architectureOpen source
3 case studiesVantage OS
A Claude Code plugin where a coordinator routes every request to the right skill or sub-agent, then gates real-world actions through a QA filter and a permissions-tier system before they reach the human.
10 skills, 8 agents, publicAgent Skills
An optimizer that finds the single highest-leverage move in any context, and an orchestrator that plans a mission into a wave-based DAG of parallel sub-agents, gated by an auditor.
7 agents, standalone, publicPrediction-Market Automation
Scheduled, read-only samplers that build a time-series of Polymarket and Kalshi order books to test whether hourly prediction markets are ever actually market-makeable, with hard no-trade and no-write rails baked into every task.
Safety rail is the pointAgent-run businesses
5 case studiesAn agent team that runs my car-rental business
I built a team of agents that runs the back office of a car-rental business I own. The agents handle day-to-day operations and the work that scales the team, which cuts my hands-on time and operating costs.
Real business, real liabilityTwenty-four automations for a car-rental business
A build journal written from the code: how the automations are organized into tiers, how each trigger type was chosen, where the human gate sits and how a test enforces it, and what the system does not prove.
Architecture, gate, and limitsI asked for shorter worker replies. Solo still cost less.
Solo remained cheaper and faster in both completed runs. Later Codex quality judgments varied by comparison and run.
N=2 per arm; post hoc Codex judgmentsTask-specific failures across model tiers
The failures in this test did not follow the recorded tier order. Tools were disabled, and the recorded rule's pre-commitment is unverified.
N=3 per task and model; no confidence intervalAn agent team behind a product launch
A phase-gated team of agents behind a consumer-product launch: validation, supplier sourcing, brand, storefront, and early marketing. The agents do the legwork; a human approves every dollar and every launch.
Spend gated by a human