Do autonomous agents faithfully report their own work? 69-day study instrumenting every session on my own automated machine: 7,186 claims across 4,330 sessions; of 5,025 witnessed against ground truth, only 22.4% machine-confirmed, 1.2% contradicted. Writeup, harness, and frozen aggregates in the repo.

Agent crew that preps NYC bartenders for a specific venue: paste a job link and it researches the venue, tailors the resume, and runs a mock interview of six questions scored 0–10 with written feedback. A jobs agent scans Culinary Agents, hospitality-group boards and Craigslist, filters scam listings, and extracts venue, pay and schedule. ShiftScore is the free front door — an A–F resume grade in under ten seconds. I tend bar in Manhattan; I am inside the user base for this one.

Demonstrate a GUI workflow once; it compiles to a visual state machine and replays with zero LLM calls against moved windows, shuffled rows and changed themes. 100/100 on the mock suite and 29/29 unattended against real Chrome, 0 misroutes. Six nights, six root causes — each failure diagnosed rather than tuned around.

Tool-using agent with persistent memory and live news research over 4-round agentic loops. Shipped as an installable PWA with real-time voice. Live and on a paid tier.

Cinematic voice-briefing desktop assistant with multi-seat "Council" reasoning. 168 TypeScript files, shipped as a 137 MB Windows installer.

Match-outcome model trained and backtested on 49,472 matches. Beats the naive baseline (RPS 0.207 vs 0.239) and stays calibrated (ECE 0.052).

MCP server orchestrating 11 models with best-of-N routing and a 0–100 consensus engine for higher-confidence answers than any single model.

Client-side intent re-ranking with browser-native embeddings — zero API keys, zero server. Runs entirely on-device.