-
-
6-layer clinical safety pipeline · 100% recall · 0 false positives
-
ClinicalMem stack · 18 MCP · 13 A2A · Federation · 1425 tests
-
21 typed runtime invariants · X25519 + Ed25519 + ChaCha20-Poly1305 · PHI never crosses
-
Verify Replay · bit-identical Q16.16 · BitNet b1.58 · 118 KB · <1 ms on $15 Pi Zero 2
Inspiration
A doctor prescribes a drug that could kill a patient. He doesn't know the patient is on a contraindicated medication from another provider. Sarah Mitchell — 67yo, four providers, no shared records — is a real-life pattern that kills 7,000+ Americans every year from preventable drug interactions. We built ClinicalMem to be the persistent memory layer those providers don't have, with FDA-grade auditability so a regulator can replay every decision a decade from now.
What it does
Persistent, auditable, contradiction-safe clinical memory for healthcare AI agents. Ingests FHIR R4 patient data, runs every drug-pair through a 6-layer safety pipeline (deterministic table → OpenEvidence → RxNorm → multi-LLM consensus → LLM synthesis → abstention), emits findings cryptographically pinned in a SHA-256 Merkle audit chain. PHI never leaves a site — only de-identified clinical knowledge crosses provider boundaries via 21 typed runtime federation invariants enforced by the MIND compiler. Curated rule packs cited inline: DDInter 2.0, AGS Beers Criteria 2023, FDA boxed warnings, CDC ACIP vaccine guidance, PubMed evidence (RCTs / systematic reviews / guidelines).
How we built it
ClinicalMem v4.1.0. Path B BitNet b1.58 ensemble: bundle A (50,949 ternary params, 118 KB, contra/major gate) cascades to bundle B (12,741 ternary params, 30 KB, moderate/serious/major specialist). 100% recall on 139-pair PCCP cohort: 44/44 contraindicated · 4/4 major · 69/69 serious · 22/22 moderate · zero false positives. Federation enforced by JointMemoryFederation.flow.mind (X25519 sealing + Ed25519 signing + ChaCha20-Poly1305 AEAD). FastMCP 2.x server with 18 tools + Google ADK A2A agent with 5 skills · 13 tools, both on Azure Container Apps. Bit-identical Q16.16 fixed-point arithmetic — the in-browser BitNet replays the same forward pass byte-for-byte, under 1 ms on a $15 Raspberry Pi Zero 2. 1425 tests passing.
Challenges we ran into
Single-model BitNets cannot fit 4-class fine discrimination in a 193-dim feature space — three retrains (v9 / v10 / v11) regressed contraindicated recall by 3–8 anchors before we settled on the disjoint-specialist cascade. Cross-architecture determinism: any non-Q16.16 op breaks bit-identical replay across x86 / ARM / CUDA. HIPAA enforcement at compile time, not runtime: the 21 federation invariants had to be expressible in MIND's type system.
Accomplishments that we're proud of
FDA-grade Verify Replay: click in browser, your machine recomputes the exact server hash. 100% recall on every severity class with zero false positives in the contraindicated class. One ternary model under 150 KB that runs in <1 ms on a $15 Raspberry Pi Zero 2. Tamper-proof SHA-256 Merkle audit chain aligned with HIPAA § 164.312(b).
What we learned
"I don't know" saves lives. The abstention gate refuses to answer when evidence is insufficient — that's a feature, not a bug, in healthcare AI.
What's next for ClinicalMem
Live integration with EHR FHIR endpoints, deterministic-table expansion beyond cardiovascular and diabetes, native vaccine pathway via CDC ACIP, and FDA SaMD pre-submission for the Layer 4.5 ternary classifier.
Log in or sign up for Devpost to join the conversation.