IX Biomedical — Agentic document intelligence
A multi-agent pipeline that reads, classifies, and extracts structure from clinical and research documents — with a human in the loop at every low-confidence step.
Architecture
The problem
IX Biomedical deals with a steady stream of clinical and research documents that don’t arrive in any consistent shape. The team was reading each one end-to-end, pulling out the fields that mattered, and cross-referencing them by hand. It was slow, and the slow part wasn’t the reading — it was the re-reading every time a document looked even slightly different from the last one.
Approach
Rather than bolt an AI step onto the existing workflow, I rebuilt the intake as an agentic pipeline. A coordinator agent takes each incoming document, decides what kind of thing it is, and hands it to the right specialist: one to classify, one to extract structured fields, one to validate the result against a reference knowledge base. Anything that comes back below a confidence threshold is routed to a human review queue instead of being silently written through.
The key decision was the human-in-the-loop gateway. The pipeline is fast because it only asks a person to look at the hard cases — the ones the agents flag — instead of every case. Sensitive material stays in-house because the whole thing runs self-hosted on hardware the client controls, with local models available for the parts that shouldn’t leave the building.
What I shipped
A working service, not a notebook. Document intake, orchestration, extraction, and the review queue all run as a deployed system the team can actually use. Every step logs its reasoning, so a wrong extraction is traceable back to the agent and the decision that produced it.
This engagement is where I learned the habit this whole site is built on: build the thing, then verify the output, then trust it. A convincing demo is not the same as a system that behaves on real, messy input.