ClinSight: Agentic Multimodal Clinical AI

Every year, 795,00 Americans are harmed by delayed diagnosis in emergency departments. Preliminary chest X-ray review takes 30-60 minutes. Rural hospitals wait 4-24 hours for teleradiology. ClinSight is an open-source multimodal clinical intelligence system built entirely on AMD hardware. It ingests chest X-ray images, lab values, vitals, and triage notes simultaneously — then reasons across all modalities through a compiled LangGraph agent pipeline on AMD Instinct MI300X via ROCm 7.0 and vLLM. Architecture: 5 parent agents orchestrate 7 subagents (12 reasoning nodes). Coordinator validates input and runs pediatric safety gates. Radiologist analyzes X-rays via Qwen2.5-VL-7B. Lab Analyst detects critical values and correlates patterns. Safety runs 3 parallel checks (contradiction, hallucination guard, bias audit) with a merge node. Documenter produces deterministic ESI scoring, differential diagnosis, and structured reports. Dual-Model Stack: Qwen2.5-VL-7B-Instruct (vision, ~14GB) + Qwen3.5-35B-A3B MoE (reasoning, ~70GB) = ~99GB / 192GB HBM3. Both models served simultaneously via vLLM on ROCm 7.0 — impossible on H100 80GB without quantization. Live Evidence: 50-case pure CXR benchmark on real MI300X. Mean latency: 23.02s. All 50 cases live, zero cache. GPU utilization: 10-49%, power 231-263W. rocm-smi evidence captured at baseline, during, and post-inference. Safety & Rigor: Physician-in-the-loop by design. Pediatric gate blocks adult-trained recommendations for under-18 patients. Bias auditor stratifies by age and sex. ESI scoring is rules-based, never LLM-generated. Apache 2.0 license.