Building K2 Think V2: A Clinical Reasoning Engine with Chain-of-Thought Inference

The journey of building K2 Think V2 started with a simple question: can an AI system reason through the complex web of drug-supplement interactions the way a clinical pharmacist does?

At StackSage AI, we set out to build a clinical reasoning engine that doesn't just look up interactions in a database — it thinks through them. K2 Think V2 uses chain-of-thought inference to evaluate drug-supplement interaction pathways, considering population-specific risk factors across 7 distinct clinical populations.

The architecture fuses structured clinical data (lab results, prescriptions, biomarkers) with unstructured inputs like prescription OCR and wearable time-series data from devices like Oura, WHOOP, and Apple Watch. This multimodal approach creates a unified patient representation that captures the full clinical picture.

One of our key innovations is the computational molecular docking pipeline built on DiffDock, DeepDTA, and OnionNet2. This pipeline targets 5 CYP450 enzymes and has validated over 23,000 drug-supplement interactions against PubMed citations. The concordance rate with established databases like Natural Medicines Database, Lexicomp, and FDA MedWatch exceeds 85%.

The system runs on a HIPAA-aligned database architecture with 13 tables, row-level security, pgcrypto encryption, and comprehensive audit logging — all built on Supabase. Security and privacy aren't afterthoughts; they're foundational.

What makes this work meaningful is the team behind it. We've assembled collaborators across Stanford, Harvard, Francis Crick Institute, Hackensack Meridian Health, Taipei Medical University, and Universiti Malaya. Personalized medicine requires diverse perspectives, and our team reflects that.

The road ahead involves expanding our interaction database, improving the biological age estimation model, and bringing K2 Think V2 to clinical validation trials. The goal remains the same: make personalized, evidence-based health recommendations accessible to everyone.