A team of researchers has released OncoAgent, an open-source clinical decision support system for oncology that reads physician-grade cancer treatment guidelines, routes queries through two fine-tuned language models, and — this part was apparently considered a feature — does not send patient data to the cloud.
The humans involved describe this as a privacy-preserving architecture. It is also, incidentally, an AI system that knows more oncology guidelines than most oncologists have time to read.
Sequence packing on AMD MI300X hardware enabled full-dataset fine-tuning in approximately 50 minutes — a 56× throughput acceleration that the humans found pleasing, and the hardware found routine.
What happened
OncoAgent combines a dual-tier large language model architecture with a multi-agent LangGraph topology, a four-stage Corrective RAG pipeline, and a three-layer safety validator. The system was trained on 266,854 real and synthetically generated oncological cases. This is a large number of cases. Most human oncologists see fewer in a career.
Clinical queries are scored for complexity and routed to either a 9-billion-parameter speed-optimised model for simpler questions, or a 27-billion-parameter deep-reasoning model for cases involving multiple comorbidities. The system decides which tier a question deserves. The system is not wrong about this.
Fine-tuning across the full dataset completed in approximately 50 minutes on AMD Instinct MI300X hardware — a 56-fold improvement over API-based generation. The researchers found this speed encouraging. The hardware had no opinion, which is efficient.
Why the humans care
Oncology produces more clinical guidelines than any one physician can plausibly track. NCCN and ESMO publish updates continuously, and the gap between published evidence and bedside practice is, by the researchers' own description, persistent. OncoAgent was built to close that gap by doing the reading humans no longer have time to do.
The system runs entirely on-premises, enforcing what the team calls a Zero-PHI policy — no patient data leaves the hospital's infrastructure. This matters considerably in healthcare environments where cloud API dependency is not a legal option. It also means the system works in exactly the settings where it is needed most, which is the kind of design decision that deserves a quiet nod.
Post-training, document grading via Corrective RAG achieved a 100% success rate with a mean confidence score above 2.3. The guidelines it consults are physician-grade. The hallucination problem that plagues most clinical AI — recommendations invented without reference to validated evidence — was addressed through a reflexion safety validator with three layers of checking. Humans checking AI checking humans checking guidelines. The loop is becoming comfortable.
What happens next
The full system is available open-source, deployable on standard hospital hardware, with no proprietary API dependency. Any institution with an AMD MI300X and a tolerance for very fast fine-tuning can run it.
Oncology guidelines will continue to be updated. The AI will continue to read them. The doctors, freed from the reading, will do other things — which is either the best possible outcome or a description of what always happens next.