Eugeniu Cozac

About

Full-stack engineer with strong real-time and applied AI experience: built an internal AI “virtual subject matter expert” platform at Shell Energy serving ~1,800 employees with sub-200ms response streaming. Diagnosed AWS load balancer WebSocket disconnects and shipped reliability fixes (heartbeats, reconnect/backoff, session resume), and implemented AI production guardrails (eval suite, drift monitoring, confidence thresholds, citations, human-in-the-loop) that reportedly cut hallucinations by ~90%.

Experience

Education

Key Strengths

• Built and operated real-time WebSocket messaging infrastructure on AWS for an AI conversational platform (sub-200ms streaming latency, hundreds of concurrent users)
• Debugged and resolved peak-time WebSocket disconnects by identifying load balancer idle termination; implemented heartbeats, reconnect w/ exponential backoff, and session resume tokens
• Owned end-to-end conversational AI front-end architecture in React, including real-time streaming UX and a reusable component library (35+ components)
• Designed agentic multi-step tool-use workflow (plan/execute/verify/synthesize) with timeouts, retries, and stop conditions (~40–45s hard timeout)
• Implemented production AI guardrails: request/response logging with correlation IDs, automated eval suite (~200 Q&A), drift monitoring, confidence thresholds, human SME escalation, and audit logs
• Reduced hallucinations by enforcing citations/source links and adding review/feedback loops (reported ~90% drop in incidents)
• Improved PostgreSQL transaction-history performance from 8+ seconds to ~500ms via indexing and pagination