I'm an AI and software engineer in Boston, MA. At Hime I build Wheelbase end to end, from Postgres security policies to the retrieval layer to the React app dealers use every day. On my own I built Grammario and Lociros, NLP products for language learners, and at the Massachusetts Executive Office for Administration and Finance I built AWS Bedrock pipelines that translate financial documents without breaking their layout.
I build the whole path a feature travels: the schema and its security policies, the retrieval and ranking layer, the model call and its guardrails, the API, and the interface someone actually touches. Then I measure it. The Go rewrite of Wheelbase's backend cut database load by more than 80%; parallelizing Grammario's analysis pipeline took a nine-second wait down to four.
My rule for AI work is structure first, model second. Grammario only lets the LLM explain grammar after a deterministic parser has already found it. Lociros never trusts a model's claim that a passage is A2; a morphological analyzer checks every word. Wheelbase's assistant can write SQL, but only inside a hardened Postgres function it cannot escape, and destructive writes wait for an approval token. Language models are powerful and unreliable, so I put them where their fluency helps and put hard checks where their mistakes would cost something.
I learn fast by building. Wheelbase ingestion got too slow in Python, so I learned Go and rewrote it . I picked up a pandas plotting bug and got the fix merged upstream. Outside of code I read history and science, and I am always partway through learning another language.
