Doctronic: Clinical Reasoning Engine & Medical Knowledge Graph Judge

Engineering Mission

Doctronic is a consumer telemedicine platform pairing an AI doctor with licensed physicians for comprehensive clinical assessment, EHR integration, and safe automated prescriptions. My role spans clinical reasoning engine development, safety guardrails, medical knowledge graph grounding, and rigorous cross-validation benchmarking.

Core Systems Built

1. Clinical Reasoning Sol Parity (doctronic-core)

• Parity with Reference Models: Engineered production Sol parity on Tier 2 clinical reasoning engines (+5,485 lines), matching Claude 3.5 Opus-high on overall accuracy (94.0% OA, 95.8% PPV) with fewer bad approvals across $n=2,037$ patient visits.
• Safety Levers & Deterministic Stops: Implemented deterministic clinical stops, override narrowing, and sourced Clinical Guidelines (CoG) enforcement.
• Leave-One-Out Safety Validation: Validated individual safety levers using leave-one-out ablations: bipolar stop (23 bad approvals saved), override narrowing (6 fixed for 0 broken), and Engine D monitoring levers (13 fixed for 3 broken).

2. Medical Knowledge Graph Grounding & Judge (fdb-graph-explorer)

• Graph-Grounded LLM Judge (ADR-034): Architected an in-process evaluation framework (GraphFacts) that queries First Databank (FDB) graph topology and Human Phenotype Ontology (HPO) plain-name findings directly in code rather than rendering flat text prompts.
• Clinical Graph Vocabulary: Implemented formal graph classification semantics (no_node, node_proposed, unlinked, linked_via, rules_fired) preventing LLM hallucinations over medical guidelines.
• Latency & Coverage Optimization: Mapped 837 of 860 Rx medication nodes to FDB ingredient synonyms, expanded acute condition families (delirium, musculoskeletal emergencies, skin infections), and accelerated test runs by ~60% (39s vs. 94s).

3. Real-World Cross-Validation & Statistical Evaluation Harness (doctronic-issue-triage)

• 5-Fold Cross-Validation Ablation Framework: Designed and executed pre-registered multi-arm cross-validation suites with paired bootstrap statistical tests (compare_arms) on AWS Bedrock.
• Structured LLM Judges: Dispatched structured triage judges with cryptographic run registration, cost ledgers, and k-majority voting with formal abstention semantics.
• Infrastructure Resilience: Handled transient Bedrock 5xx exceptions with exponential backoff and automated instance-profile EC2 run hosts.