Emiliano Mankolli

GenAI & RAG System Development

A build follows the same arc every time. I start by analyzing the problem: what questions the system must answer, who asks them, and what a wrong answer costs. From there I design the solution and the system around it. Then comes the knowledge base itself: processing your documents so retrieval works on your actual content, since a corpus of legal decisions behaves nothing like a folder of engineering manuals. With that in place I build the pipelines that connect ingestion, retrieval, and generation, then move into validation and, once live, ongoing maintenance.

Accuracy gets tested two ways. Before launch, I run back testing: questions with known correct answers, checked against what the system returns. After launch, evaluation continues on production traffic, because real users always ask things a test set did not anticipate. Grounding every answer in retrieved documents is what keeps the system honest. When it drifts, back testing shows where.

Production ready means exactly that distinction. A demo answers well in a controlled session. A production system keeps answering well next month, on live traffic, with monitoring and a maintenance path when the knowledge base grows or the failure modes shift.

What I need from you is simple: access to the data, and one person who knows your domain deeply. That domain expert is the most important ingredient. They tell me which answers are actually correct, which is something no model and no engineer can decide alone.

For a sense of what this looks like at scale, the work below includes a system built on a corpus of more than one million documents.

Work together