manuel jose

I build and evaluate AI systems end to end, from a 52.9M-parameter transformer trained from scratch to production GraphRAG and multi-tenant SaaS products.

For Sol, I trained a decoder-only transformer without pretrained weights, identified and removed 6,304 cross-split duplicate documents, and built 154 tests with a CI drift guard. I also built Elephant in the Room, a procurement-risk pipeline analyzing 509,358 Colombian health contracts and improving post-adjudication recall from 60% to 100% across 11 injected fraud patterns.

With Lante, I built a tenant-scoped GraphRAG platform for WhatsApp and PDF ingestion, combining Qdrant vector retrieval with FalkorDB knowledge-graph traversal and evidence-based responses. I was also the sole developer behind live products including Bodega Boost and Iter AI Parking, delivering secure multi-tenant workflows, billing, offline functionality, and AI-assisted features.

My cGMP validation work for clients including AbbVie and Fresenius Kabi shaped how I approach leakage control, traceable evidence, risk-based decisions, and measured baselines. I'm completing an MSE in Artificial Intelligence, Machine Learning and Data Science at the University of Pennsylvania.