PHAROS AI code retrieval infrastructure · 2025 to now
Coding agents need context tied to the version of the repository they are working on. I built PHAROS: Tree-sitter indexing of symbols, imports and calls, retrieval filtered to one repository, and results served over REST and native MCP. The decision that shapes it: returned code is checked against the Git blob it was indexed from, instead of trusting a search result by default. That proves integrity, not correctness.
◐ Reported Held-out nDCG@10 rose 0.394 → 0.697 and 0.463 → 0.726, across two held-out sets totaling 101 questions; single annotator, not reproduced here. Repository memory v1 failed its primary gates: 55% audited precision, primary source-file recall 0.827 → 0.823. Beside that failure, file-selection precision rose 63.4% → 77.5% in an 80-task test ◆ Owner confirmed.
How it picks a search mode
Auto mode prefers dense retrieval, uses hybrid with hints and falls back to keyword with a warning. An explicitly requested mode that is unavailable fails. Sparse (SPLADE) retrieval is implemented in an earlier search module; the audited auto mode does not route to it.
Noisy showroom conversations had to become coaching feedback grounded in what was said and in versioned product specifications. In an EV showroom coaching pilot I benchmarked on-device speech processing, built analysis that links each finding to transcript evidence, and built the evaluation. The decision: code with numeric tolerance bounds decides whether a specification claim is correct; the model interprets the conversation.
◆ Owner confirmed 96.5% speaker attribution on 50 hand-annotated recordings. A 100-session evaluation harness. Separately, a 40-session A/B test: 24% fewer prompt tokens, 11 of 11 planted errors still caught. Résumé results, not reproduced here.
Valerie / CognitiveOS Native AI assistant with validated actions · The Keeper’s Study
CognitiveOS is my native macOS and iOS app, and Valerie is its assistant. A model proposes one typed move; deterministic code checks it against the state it was issued and the bounds of its arguments before anything runs. This site’s guide uses the same kind of boundary, acting only on these pages. Try to break it.
Magna International Software Intern · warehouse robotics · summer 2025
An operator needed to be able to step in while a warehouse robot moved. I built the React interface (camera feeds, keyboard and joystick) and the Python/FastAPI backend that sends motion commands over SSH, with operator-initiated hand-offs between autonomous and manual control. Command latency: 58 ms keyboard, 62 ms joystick ◆ Owner confirmed.