E4L-Prometheus

E4L-Prometheus generates clinical treatment plans for Parkinson's patients by combining data from an EHR (Cerbo), a CRM (GoHighLevel), and lab results, then running it through an LLM pipeline with a provider-facing review and feedback workflow.

Led the platform-wide LLM migration from OpenAI to Anthropic Claude on AWS Bedrock: designed a provider-abstraction layer (text, JSON, vision, and structured-output APIs) that became the single integration point for all AI calls, enabling environment-driven provider switching and safe model upgrades — a 119-file refactor.

Architected a supplement recommendation engine using RAG over AWS Bedrock Knowledge Bases — clinical rules extracted from practitioner transcripts, deterministic pre-LLM selection over a 500+ product catalog with fuzzy matching. It's the platform's largest subsystem, backed by 112 unit tests.

Designed an asynchronous clinician-feedback pipeline (DynamoDB Streams → Lambda → SQS FIFO → LangGraph workers) that classifies provider corrections to AI-generated plans and proposes fixes, with conflict detection and an insights analytics API.

Built the nightly batch plan-generation service on ECS Fargate with EventBridge scheduling, two-level deduplication, and per-plan token/cost tracking with CloudWatch alarms. Hardened HIPAA compliance across the stack: PHI scrubbing in structured logs, customer-managed KMS encryption, and immutable audit archives.

Resolved production reliability issues including subprocess-isolated PDF rendering to survive native crashes in nightly batches, and a PDF-processing migration that cut document-processing time roughly 17x.

Designed and built the provider-facing web app in Next.js, TypeScript, and Tailwind CSS: a feedback-management dashboard, AI-insights analytics, scheduled-plan review with one-click EMR write-back, and a WYSIWYG treatment-plan editor with locked clinical sections and version history — authoring roughly 30 of the app's 43 components.