Mario M.

Most LLM products work on the demo. Few survive production. I build the layer that keeps them honest: 8-tool RAG agentic chatbot with citation grounding, OCR+Gemini document extraction verified on 20 cases / 65-case batches, all on FastAPI. I specialize in production LLM systems where the data is messy and the answers must be grounded: agentic RAG for domain queries, document intelligence (OCR + entity extraction + enrichment), and Python automation that connects AI outputs into real business workflows. End-to-end ownership: ingestion, retrieval, LLM orchestration, evaluation, observability, deployment. 🛠 Recent production work • Agentic AI chatbot for a US construction-tech startup. Built the v3 backend: 8 RAG-grounded tools in a streaming agentic loop (max 12 rounds per turn), Groq LLM at temperature 0 for deterministic outputs, citation-grounded responses, anti-hallucination guards (asset-type mismatch detection), and a 9-milestone migration with a legacy-vs-v3 parity audit I authored. • Document AI pipeline for a Florida real-estate firm. Court-portal Selenium automation + PDF OCR + Gemini entity extraction → up to 7 named parties per case (name, address, relationship, DOB), verified end-to-end on 20 cases and configured for batches of 65+. Multi-threaded enrichment, Google Drive / Sheets integration. 🧠 Where I'm different Most engineers stop at "it works once." I build systems that stay working: paraphrase-stability validation, regression tests on LLM outputs, voting consensus across paraphrased queries, observability for tool-call traces, and prompt-update detection so silent regressions surface before clients notice. 🔧 Core stack Python • FastAPI • Pydantic-AI • LangChain • OpenAI • Anthropic • Gemini • Groq • Tesseract • Poppler • Vector DBs • Redis • Docker • AsyncIO • Selenium • n8n • SQL 📊 Adjacent capability When the project needs it: classical ML (XGBoost, Random Forest, clustering, NLP), production data pipelines, Python automation that connects AI outputs into real business workflows (CRMs, Sheets, messaging, webhooks). If you need production AI on messy real-world data — grounded, observable, reliable — let's talk.