Know Your Trial
Stop manually searching for clinical trials.
Created on 26th July 2026
•
Know Your Trial
Stop manually searching for clinical trials.
The problem Know Your Trial solves
Finding clinical trials for rare diseases (like Fabry, Pompe, or Gaucher disease) is notoriously difficult and heartbreakingly slow. Clinical trial coordinators and patients often spend hours manually cross-referencing dense, unstructured patient histories (PDFs) against highly technical, jargon-heavy inclusion/exclusion criteria on ClinicalTrials.gov.
The biggest bottleneck? The semantic gap. Traditional keyword-matching algorithms completely fail in healthcare. If a trial protocol requires "chronic nerve pain", but a patient's medical note states "severe distal paresthesia", standard database queries will miss the connection entirely—potentially costing a patient access to a life-saving experimental therapy.
"Know Your Trial" is a multi-agent AI pipeline designed to eliminate this bottleneck. We built a deterministic intelligence engine that automates the entire matching pipeline:
Automated Ingestion: It parses massive, unstructured patient clinical PDFs (simulated via our GPT-2 Data Gen).
Semantic Extraction: A specialized BioBERT NER Agent extracts phenotypes, labs, and medications, mapping them to standard Human Phenotype Ontology (HPO) terms to bridge the vocabulary gap.
High-Speed Retrieval: It queries these embeddings against active clinical trial protocols stored in our RAG Pipeline (Supabase + ChromaDB).
Explainable AI: The Orchestrator Agent delegates the final scoring to our Explainability Agent (Gemini 2.0 Flash Lite), which generates a physician-level breakdown of exactly why a patient passed or failed.
What used to take a medical coordinator 4 to 6 hours per patient is now executed with higher precision in under 15 seconds. It doesn't just match patients; it democratizes access to experimental treatments by making clinical trial discovery instant and universally accessible.
Challenges we ran into
-
The "Black Box" Trust Barrier in Healthcare Early in development, we realized that generating a "98% Match Score" was useless. In clinical environments, doctors do not trust opaque AI decisions. If the system couldn't justify its reasoning, it would never be adopted. How we got over it: We pivoted our architecture to prioritize transparency. Instead of using a single LLM to guess the answer, we engineered an Explainability Agent powered by Gemini 2.0 Flash Lite. We forced the pipeline to return deterministic, citation-backed JSON payloads. Now, the UI renders a line-by-line breakdown report, citing the exact sentence in the patient's record that satisfies the specific NCT protocol criterion. We turned a black box into an auditable clinical tool.
-
Hallucinating Temporal Constraints As we scaled the matching engine, we hit a massive hurdle with temporal logic. Our early semantic search was incredible at mapping phenotypes (e.g., matching "enlarged spleen" to "splenomegaly"), but it completely hallucinated strict temporal boundaries. For instance, a Gaucher disease trial excluded patients who had been on Enzyme Replacement Therapy (ERT) for less than 12 months. The LLM would see the keyword "ERT" in the patient file and flag them as eligible, completely ignoring the timeline. How we got over it: We realized you can't use language models for discrete math. We ripped out the pure-LLM matching approach and built a Hybrid Multi-Modal Engine. We introduced a deterministic Python scoring layer utilizing Regex to handle numeric, categorical, and temporal data (like calculating exact months between dates). We then built an Orchestrator Agent that acts as a router—sending temporal rules to the Python engine, and sending complex linguistic rules to the BioBERT NER Agent. By decoupling the logic, we eliminated temporal hallucinations entirely while maintaining deep semantic understanding.
Tracks Applied (2)
Best Use of Gemini API
Major League Hacking

