Multi-model pipelines
Each model in the role where it’s best — OCR, long-context verification, recovery — chosen empirically, not by vendor preference. Measured, reproducible.
No demos. Production LLM systems — one used by a paying law firm, one by a state government. I build the pipeline, the backend, the frontend and the deploy. Plus the part most people skip: making the model’s output reliable enough to act on.
Building with AI isn’t the hard part anymore. Anyone can wire up an LLM demo. The hard part is a system that survives production.
That verifies. That doesn’t hallucinate where it counts. That a real client bets their money — or their legal rights — on.
That’s what I build: not the prototype that wows on a call, but the system still standing six months later.
Each model in the role where it’s best — OCR, long-context verification, recovery — chosen empirically, not by vendor preference. Measured, reproducible.
Agents that work in parallel, coordinate, and recover what a single pass leaves behind. The same architecture that runs this operation.
The part most people skip: automated quality gates so the model’s output is reliable where an error costs dearly. Legal, government, money.
Real systems, in production, with users who pay or govern with them.
Multi-model pipeline (Claude / Gemini / Groq) verifying 1M+ legal records with automated quality gates. In production, paying client.
02/03Multi-source geospatial ETL over 645 municipalities × 104 variables. Dataset published with a DOI, validated in a state-government workshop.
The agent system that runs my own operation: persistent memory, domain skills, multi-agent orchestration. The same head I direct for a client.
Python · FastAPI · Next.js · TypeScript · Claude / Gemini / Groq · PostgreSQL · Docker · CI/CD — from the data pipeline to the deploy. I work end-to-end: I don’t ship one piece and leave the rest to someone else.
Anyone can wire up a demo.ARLEX PERALTA — AI · AGENT ENGINEER
Surviving production, not so much.
Remote hire or a one-off agent-automation project. Tell me what you need to build and I’ll tell you if I’m the right person — and how I’d do it.