

TrialNexus is an AI-native clinical trial recruitment platform built to solve one of medicine's hardest problems: finding the right patients for the right trials, fast. Pharmaceutical companies lose millions every month when trials recruit slowly. Clinicians spend hours manually reviewing patient records. TrialNexus automates both sides. At its core, TrialNexus uses Google Gemini multimodal embeddings to encode both MRI brain scan images and clinical text together into a single vector, enabling semantic patient search that understands medical context — not just keywords. Patients are ranked by match percentage against trial inclusion/exclusion criteria. Phase 2 Lab Screening goes deeper: an AI agent runs each candidate through a structured SQL-based eligibility gate, automatically disqualifying patients who fail criteria and explaining exactly why — with full lab detail visible on click. The Commercial Intelligence module runs a 4-agent pipeline using Server-Sent Events for real-time streaming. It synthesizes competitive landscape data, KOL insights, site feasibility, and market opportunity into a downloadable report — powered by Featherless.ai's Qwen2.5-72B model. The entire stack is self-hosted: FastAPI backend, React 18 frontend, and Elasticsearch 8 running on Vultr — fully deployed via Docker Compose on Coolify with zero cloud lock-in. TrialNexus turns weeks of manual patient screening into seconds of AI-assisted matching.
19 May 2026

Cardiology ICUs run on a painful bottleneck: the senior cardiologist spends most of the day in surgery, while nurses and junior doctors at the bedside wait for them on every decision. Patients wait. Care is delayed. AMD Cardio Agent removes that wait time. A nurse enters the patient's current state and vitals, then uploads cardiac imaging — 12-lead ECG, chest X-ray, echocardiogram, and cardiac CT. The agent bundles that with simulated EHR data including labs, patient history, and a 6-hour vitals trend. It then runs Qwen2.5-VL-7B on an AMD MI300 GPU via vLLM and produces a structured preliminary diagnosis with recommended bedside actions and key supporting evidence. The junior doctor reviews everything on a visual board: all uploaded imaging displayed side-by-side, a Plotly vitals trend chart over the last six hours, a cardiac biomarker chart with abnormal values flagged against reference limits, an accordion of patient history, and the AI's draft diagnosis. The junior approves it as-is or edits it. Both actions feed a sentence-transformer RAG knowledge base — approvals save as confirmations that validate the pattern, edits save as corrections that override it. The next similar case retrieves these and injects them into the prompt, so the agent gets smarter every shift, learning from both juniors and seniors with no fine-tuning required. The frontend is a Gradio Space on Hugging Face. Inference runs entirely on the open-source AMD stack — Qwen2.5-VL-7B-Instruct, vLLM, ROCm, MI300 — fully self-hosted, with no cloud dependency for the model itself. The architecture mirrors how a real Dutch cardiology team would deploy this in production: imaging, labs, and history flow in; a structured draft flows out; every doctor interaction trains the system.
10 May 2026