AI systems.
Tested in public.
I build agentic workflows with a full-stack development background: bounded context, validated outputs, controlled execution, and evaluations that expose the limits.
Primary showcases
Code, controls, and outcomesFrontend Director
AI-powered React editor with bounded context, validated edits, user approval, and reversible history. Structural additions separate read-only planning from approved execution.
- Context
- Selected JSX / bounded source
- Controls
- Patch validation / stale-plan rejection
- Evidence
- 125 automated tests / atomic history
BTC–Jev
A deployed AI decision experiment with scheduled market processing, deterministic guardrails, limited model requests, and a public evaluation dashboard. Paper-only; no real orders.
- Decisions
- Candidate gating / long-only paper policy
- Controls
- Daily request reservations / pause switch
- Evaluation
- Simulated fills, costs, and outcomes
Problem Hunter
An AI research pipeline with source-backed findings, structured evaluations, caching, cost controls, and audit reports. Findings need human review; reliable autonomous opportunity discovery is not established.
- Evidence
- Attributed source passages / dates
- Controls
- Request ceilings / per-run cost limits
- Outputs
- SQLite ledger / JSON and Markdown
Frontend Director is a development prototype without a build-before-apply gate. BTC–Jev keeps archived direction forecasts separate from current paper results. Problem Hunter's general remedy research requires a reviewed workflow brief.