Featured
Five public repos, one costed case study, one production system
Script in, captioned video out, cost logged
Seven-stage video pipeline in Node and ffmpeg, built on the ElevenLabs stack. Feed it a plain-text script, out comes a finished, captioned short. Every API call logs to a costed run manifest. Each run is auditable down to the cent.
Scores drafts against a target voice
Built a Python pipeline on Claude that scores drafts against a target voice on six stylistic axes and blocks anything below a QA threshold. "Sounds like me" isn't a spec, so this makes it one. Calibrated on 6.9M+ words of executive comms, with dual-persona review and a deterministic offline mode. Method and sample corpus are in the repo.
Job-search automation, in production
A job-search automation system: ATS portal scanning, LLM role evaluation, CV generation, and batch processing. Runs as a production fork of open-source santifer/career-ops. Orchestrates multiple models across roughly 50 scheduled agents on macOS launchd. Powered 740+ role evaluations. Operated software, not a demo.
A tax checker that must cite or stay silent
A Claude-based agent that checks tax filings against a four-layer knowledge base: federal code, state statutes, IRS publications, precedents. It refuses to answer without citations: a confident wrong answer about taxes is worse than none. In practice it caught a roughly $19K error that commercial filing software had pre-filled and defended.
Triage agent for internal comms requests
A sanitized public demo of an internal-communications triage agent, built on Google Apps Script and Gemini. It classifies incoming requests, drafts revisions, and escalates the rest through three sequential prompts with conditional knowledge-base loading. A confidence threshold routes anything the model is unsure about to a human instead of letting it guess.
Budget-gated wardrobe recommendations
A TypeScript PWA that recommends wardrobe purchases under a real budget. The LLM only pulls product data from a page. Three deterministic code gates (budget, climate, aesthetic rules) issue the verdict, and unknown data returns INSUFFICIENT_DATA instead of a guess. An append-only audit log keeps the whole decision path inspectable.
