I build the data products
that make AI work —
and the AI itself.

Flagship · Designed, built & deployed solo

Job Tracker — a scheduled, agentic document pipeline

Next.js 16 · Hono · Cloudflare Workers + D1 · Claude (headless, MCP) · Python

The idea

Most "AI document" tools are a chat box and a prompt. I wanted to answer a harder question: can an LLM run unattended, on a schedule, and produce work you'd actually trust? I picked a problem I had real stakes in — my own job search — and built a system that, every night, turns a structured profile plus a job posting into tailored application documents, to a quality bar I define, while showing its reasoning so I stay in control. The job-search use case kept me honest about quality; the agent architecture is what transfers to any high-stakes, repeatable knowledge work.

Decisions I'm proud of

• LLM as orchestrator, not a chatbot. A runbook drives a multi-step, tool-using agent loop — ingest, dedup, reason, generate, render, publish, report.
• A hard line between plumbing and judgment. State transitions, rendering, auth, and uploads are deterministic code (zero AI); only the genuine judgment calls go to the model. Knowing which is which is half the design.
• Human-in-the-loop, by construction. The agent writes a reasoning log every run — what it emphasized, stretched, and skipped — so review is an approval step, not blind trust.
• Guardrails as a contract. Non-negotiable content rules plus a structured output schema keep every autonomous run on-spec — the start of a real eval loop.