PRASHANTH BIYYANI

Summary

Technical Skills

Experience

Own ELI end-to-end — a production conversational-AI assistant deployed across Florida's medical-board sites

• Own ELI end-to-end — idea, design, deployment, and maintenance — the department's conversational-AI assistant embedded across Florida's medical-board public sites (flhealthsource.gov/boards); built the backend and orchestration in Node.js/Express (resilient REST APIs, custom async middleware) serving 100,000+ monthly users.
• Sustained 99.9%+ uptime on Kubernetes (IBM Cloud) by owning rollout strategy, resource-limit tuning, IAM policy, and monitoring for the ELI platform.
• Built Python ETL and analytics pipelines processing ~1–2M records/month (10–20 per user across 100K+ users) into a downstream PostgreSQL store, powering near-real-time operational dashboards for department leadership.
• Built and maintained the ELI chatbot and Conversational Insights front-ends in Angular/React, integrating with backend services over REST for a responsive real-time experience.
• Harden security with IBM Cloud authentication, IAM policies, and access controls to protect sensitive health data and support compliance.

Backend microservices & full-stack (MEAN) for healthcare applications

• Built cloud microservices in Java/Spring Boot (Spring MVC, centralized config server, client-side load balancing), centralizing configuration to speed new-service onboarding.
• Developed MEAN-stack applications (Express, Angular 7, Node.js) with reusable components, RxJS Observables, and @NgRx state management for complex multi-step forms.
• Designed REST APIs (Node/Express) serving JSON to Angular front-ends and integrated user-facing components with server-side services.

Selected Projects

• Governed Clinical-Analytics Copilot (solo) — Self-hostable NL→SQL clinical-analytics agent over EHR data. Python, PyTorch, sqlglot, LangGraph. GitHub
  • Designed a 5-layer AST-based SQL guardrail stack (sqlglot) blocking destructive queries, PHI columns, and small-cell (k-anonymity) results.
  • Fine-tuned a 7B code model with QLoRA + ORPO for NL→SQL, with learned abstention on unanswerable questions.
  • Built a 1,200-line evaluation harness measuring execution accuracy and answer/abstain reliability.
• ChainWatch — Supply-Chain Risk Monitor (co-built) — Owned the data ingestion (7 live sources: GDELT, NOAA, USGS, FRED, …), LLM-cost controls (keyword pre-filtering, request batching), and the 10-tab dashboard. Python, Gemini, Streamlit. GitHub
• Building-Damage Segmentation (solo) — PyTorch semantic segmentation (attention U-Net with CBAM/ASPP) on 11-class UAV disaster imagery; compound CE/Focal/Dice loss for class imbalance. PyTorch. GitHub

Hackathons & Competitions

• HackHydra 2026 — Track 03 (Memory & Context Retrieval): built MedMemGraph, graph-native longitudinal memory for clinical AI agents — versioned fact graph, path-traversal retrieval, structural abstention. GitHub
• Google 5-Day AI Agents Intensive 2026 — Capstone: Cash-Flow Copilot (co-built) — agentic CFO assistant with a deterministic numeric-provenance guard. ADK, MCP, Gemini.
• Mayo Clinic Datathon 2025 — predicting progression to Berlin-defined ARDS (acute respiratory distress syndrome) for ICU care; multi-model approach (XGBoost + NEAT evolutionary neural nets + regression benchmark) balancing accuracy with clinical interpretability.
• Hacklytics 2025 — Georgia Tech — MRI 1.5T→3T cross-field-strength medical imaging (co-built). GitHub

Education

Certifications