Franco Ugarte — AI Solutions ArchitectI build applied AI that turns data and broken processes into decisions worth acting on.

The ERPs were too large for any single context window

Every question meant manual analysis across ERPs and CRMs that were never built to answer questions directly, and finding the right table before running a query was its own research project.

A semantic layer, then hybrid retrieval

I built AI that finds the right data on its own and checks its own work before answering, so decisions that used to need a specialist now just need a question.

I set the LLM and hybrid-RAG strategy: a semantic layer over every table, column, and business term; hybrid retrieval combining vector search, BM25, and reranking for schema selection; and a LangGraph validate-and-retry loop that took decision accuracy from about 60% naive to over 85%. Read-only allow-lists, row-level security, and PII masking kept it safe, with evals running as a CI gate before anything shipped. I led the 5-person technical team through all of it.

Reports that took five minutes

Slow reports and manual processes were the daily cost. Every decision waited on a query, and every query waited on infrastructure that had never been re-modeled for the volume it was now carrying.

A data model rebuilt for scale, then predictive automation

I rebuilt the data behind their reports so what used to take five minutes now takes six seconds, then layered automation on top that predicts outcomes instead of just reporting them.

I re-modeled the data architecture and tuned DAX for up to 50× faster reports, cutting time-to-insight from days to minutes. On top of that: OAuth 2.0, automated testing, code reviews, CI/CD, and load balancing for high-volume queries, plus the data model and ETL behind a claims-and-case automation platform. Predictive automation saved clients 600+ hours a year, with over 30% efficiency gains in early pilots.