Saad Azis

Built a production-grade multi-agent orchestration system for a Self-Learning Clinical Intelligence workflow, automating SOAP note generation, drug interaction analysis, and patient communication. Developed multi-tenant RAG and knowledge management for 20+ healthcare clinics with strict tenant isolation and persistent, ACID-compliant PostgreSQL-backed memory. Implemented reflective agent loops to detect tool-call/LLM failures, automatically retry with corrected parameters, and safely scale down after exhaustion. Engineered patient context storage with deep-merge updates, reduced API payload sizes by ~70%, and integrated UID-based suggestion tracking to support iterative model refinement. Improved retrieval and embedding quality using LangSmith evaluation pipelines, prompt/version control, and monitoring in production.

Developed a production-ready, HIPAA-compliant intelligent doctor assistant with voice integration and real-time interactions via text, voice, and WebSocket channels for 50+ physicians. Implemented a multi-tool agent connected to EHR data (appointments, prescriptions, lab results, and patient records) with reflective tool-call recovery for transient API failures. Added mandatory confirmation workflows for high-risk operations like appointment booking/cancellation using template-based validation, reducing medical errors by ~40%. Built temporal-aware prompt engineering for accurate date/time injection and appointment filtering. Monitored latency, tool-call success rate, and conversation quality using LangSmith dashboards and evaluation datasets.

Built a document-grounded RAG chatbot that answers using source text chunks and supports references across multiple file types. Implemented AWS Lambda ingestion that processes files uploaded to S3 for zero-touch document ingestion. Added real-time voice interactions using WebSocket for hands-free user experience.