🤓 Latest Submissions
Triage Relay
Most hospitals already have a check-in kiosk, and behind it sits a rule tree. Every branch forces a certain answer even when the evidence only leans one way, so the doubt is thrown away before the receiving department sees it, and every change needs a software engineer. Large language models are the other option, but they can't show why they chose, and they can sound sure when they aren't. Triage Relay is neither. A patient types what's wrong in their own words. Two tiny transformers (6.5M parameters each) do what language models are good at: turning everyday words into structure, a checklist of symptoms. They never decide. Counted, readable seesaws do the routing: every answer pushes toward a department by an amount counted from examples, danger first, with every reason on screen. When the evidence only leans, it says so and hands the patient to a person. Measured on held-out synthetic patients: 90.5% routed automatically, 99.89% of those to the right department; the rest go to a person, with reasons. Against a decision tree trained on the same answers, the seesaws are more accurate (98.3% vs 97.4%), and the 1 patient in 8 they flag as unsure holds nearly half of the tree's mistakes. A working app: the patient checks what was understood before the ticket; NHS red flags go straight to staff; injuries go to a nurse; NEWS2 and qSOFA for the nurse. Nurses correct routes on the spot, with no engineer and no retraining, and it never learns on its own. It runs on the hospital's own computer, sends nothing to any AI company and never trains on patients. The whole chain runs on an AMD Ryzen 7 5700X3D at 68.7 ms per patient, no graphics card. We also built a full MI300X run on ROCm (both models plus Gemma 3 270M). Our AMD cloud account was set up days earlier, but GPU use needed a separate approval, only shown at launch, that did not arrive in time. Open source (MIT, NEXICOG Ltd). Trained only on synthetic DDXPlus patients. Not a medical device; it never diagnoses.
18 Oct 2026
