Open source. Edge deployed. Safety findings published.
EdgeCDSS is an AI-assisted clinical reference for medics working far from hospitals — environments with no reliable internet, power, or backup. The entire system runs on a single small computer (an NVIDIA Jetson) that travels with the team. Every answer is grounded in established trauma care guidelines, and medication math and safety checks are handled by conventional software rather than AI — the AI's only job is understanding your question and explaining the answer. Version 4.3 is live now for research and evaluation.
Clinical AI is built for hospitals — reliable power, fast networks, specialist backup. Step outside those walls and it stops working. Conflict medicine, wilderness rescue, expedition and remote care, low-resource hospitals — the environments where decision support could matter most are the ones it isn't designed for.
Existing tools assume hospital infrastructure. We design for the absence of it. No reliable power. No internet. No physician backup. A single provider who is the entire system.
We are an open-source research initiative building the tools, methodology, and evidence base for AI in austere medicine. Built by clinicians. Evaluated by working providers. Published openly so anyone can build on it.
We document what fails, not just what works — because in clinical AI, understanding failure modes matters more than reporting accuracy scores.
EdgeCDSS is one platform, not separate products — a self-hosted edge system that uses cloud inference when the link allows and local inference when it doesn't. Three capability layers at different stages of maturity, developed in the open as a living roadmap others can build on, adapt, and take further.
Hybrid routing is only as good as the link it can measure. Alongside EdgeCDSS we build and open-source small, dependency-light utilities for characterizing the network a deployment actually has — not the one the coverage map promises.
The knowledge base, retrieval engine, deterministic safety gates, patient context and vitals tracking, the ventilator card engine, web interface, feedback system, and audit logs run on an NVIDIA Jetson Orin Nano.
Deterministic Python code performs medication calculations, contraindication checks, routing, patient-context handling, vitals capture, and safety gates. AI is used for language processing, retrieval support, response generation, and semantic validation.
Version 4.3 adds a third source of truth alongside the guideline corpus and the labelled general-reference fallback: a clinician-authored card, served verbatim, carrying its own references and a dated signing role. A card answer is neither retrieved nor generated — and the engine refuses to serve a card nobody has signed, with no debug flag or override that can reach that gate.
EdgeCDSS Version 4.3 is network agnostic. The Jetson deployment may connect through Starlink, broadband, Wi-Fi, Ethernet, LTE/5G, or another supported IP network. Remote access during development and testing used Cloudflare Tunnel with outbound-only HTTPS connections.
The AI in Austere Medicine Project is committed to developing AI systems responsibly, transparently, and with respect for privacy and human oversight — guided by established principles from medicine, cybersecurity, AI governance, privacy, and open-source software.
In field evaluation, external clinical testers identified patient safety failures that our automated test suite completely missed. We published all of them. The dosing errors, the missing surgical airway, the dangerous oxygen recommendation — documented, analyzed, and fixed in the open.
This is how clinical AI gets safer. Not by hiding failure modes behind NDAs and closed evaluation processes. By building in public, testing with real providers, and publishing exactly what went wrong and exactly how it was fixed.
Since Version 4.2 that work has a harness behind it. A standing bank of 160 scenarios — real queries pulled from session logs, prior safety findings replayed with their turn sequences, and authored edge cases — runs against a pinned server snapshot, and the before/after numbers are published with the fix. It is how “the ventilator answers feel thin” became “0 of 4 DKA phrasings return any ventilator setting, 100% reproducible, against 4 of 4 for TBI” — an opinion nobody can act on, turned into a defect with a control group and a definition of done. That measurement is the reason Version 4.3 exists.
Studies, benchmarks, and frameworks shaping AI in austere medicine — updated weekly. Each entry notes how it connects to what we're building. External links, no paywalls where possible.
New versions ship. Field data gets published. Delivered via the AI-AMP Substack — one email per milestone.
No spam. One email per major milestone. Unsubscribe anytime.