Skip to content
← all work

RAG + agent + fine-tuned extractor

A from-scratch MLOps capstone: a retrieval-augmented agent backed by an MCP server and a LoRA-fine-tuned extraction model, with an evaluation harness.

in build
status
harness
eval
PyTorch LoRA RAG MCP Python

The goal

A deliberate stretch project to build depth in the areas I’m growing into: transformer internals, LLM fine-tuning, and end-to-end MLOps. Rather than a toy demo, it’s structured like something that would ship.

What it includes

A retrieval-augmented generation pipeline, an agent that calls tools through a Model Context Protocol (MCP) server, and a domain extraction model fine-tuned with LoRA. Critically, it’s paired with an evaluation harness so that “better” is a measured claim, not a vibe.

Why it’s here

It’s honest about being in progress. I’d rather show the work taking shape — with its eval harness and structure visible — than wait to present a finished artifact. A write-up of the build is on the blog.