I build AI tools around one belief: information that only experts can read should be usable by everyone.

Vehicle recall notices, spec sheets, earnings reports, SEC filings — all public, all technically "available," and all written so that the people they matter to most can't act on them. I think that gap is one of the most fixable unfairnesses around, and that LLMs — used carefully — are the right tool to close it.

Every project below was designed, built, and evaluated by me, end to end, on my own time. They exist because I wanted them to exist.

I started with cars because the stakes are concrete — safety, money, and decisions people make once every few years without expert help.

• recallclear — translates NHTSA vehicle safety recalls into plain language a driver can act on. A LoRA fine-tune of SmolLM2-135M trained on 11,591 real recall notices.
• autotriage-ai — "Is my car problem a safety issue?" Reads a plain-language complaint, maps it to one of 14 vehicle systems, and tells you whether it's worth reporting. Trained on real NHTSA owner complaints; live demo on Cloud Run.
• carspec-ai — multi-attribute vehicle recognition from images (computer vision).
• car-review-sentiment — DistilBERT sentiment analysis over real automotive reviews, built evaluation-first.
• carrec-ai — LLM-powered car recommendations with Responsible AI constraints: grounded picks, refusal and fairness guardrails.

Same belief, higher stakes. Cheap stocks are easy to find; what ordinary investors can't do is tell a wrongly-punished company from a dying one.

• value-trap-detector (in progress) — classifies beaten-down US stocks into four categories (wrongly killed / momentum pullback / turnaround bet / value trap) and vetoes any pick it can't write a falsifiable exit condition for. In my own six-week scan of 30+ post-earnings crashes, only ~5% were genuine mispricings — so the honest output of this tool is mostly "don't buy, and here's why."
• trading_agent — an earlier build: an autonomous LLM trading agent with a hand-rolled ReAct loop (no agent frameworks) and 7 tools over the Alpaca API.

Nothing here is investment advice. These are engineering projects about reading disclosures — not stock tips.

• Numbers are computed in code. Narratives are LLM-assisted. LLMs in my pipelines never do arithmetic — they read, judge, and classify. Financial figures come from primary sources like SEC EDGAR XBRL, never from a model's memory.
• Evaluation first. A model without a labeled evaluation set is a demo, not a tool. I build the eval set before I build the feature.
• Provenance everywhere. Every judgment cites the document it came from.
• Honest base rates. If a tool says no 95% of the time, that's not a bug — refusing confidently is rarer, and more useful, than recommending confidently.