AI Engineer with 13+ years of total experience — 3 years in AI and 10 years in enterprise software engineering. Building enterprise-grade AI systems with RAG, Agentic AI, evaluation frameworks, observability, and cloud deployment.
Two flagship projects across two clouds and two architectural patterns — Enterprise RAG on AWS and multi-agent clinical intelligence on GCP.
Production-deployed enterprise RAG system for Kubernetes IT operations. Hybrid retrieval with cross-encoder reranking, CRAG, HyDE, Self-RAG reflection, Text-to-SQL with human-in-the-loop approval, 9-layer guardrails, sqlglot safety validation, semantic caching, and admin observability — all in one production workflow. Auto/Manual experience modes with preset query library.
Supervisor-orchestrated system of 6 specialist agents that autonomously audit 400K+ ClinicalTrials.gov studies for broken promises, missing results, sponsor credibility, cross-study patterns, safety gaps, and timeline deviations. Three-tier LangMem memory (Episodic · Procedural · Semantic) persists knowledge across sessions — human rejections permanently update procedural memory so agents learn from every review cycle. Per-agent HITL confidence thresholds (0.55–0.70) gate low-confidence signals to a human review queue with a learning loop that injects new rules back into agent context. Deployed serverless on GCP Cloud Run.
50-class image classifier built from scratch in PyTorch — no pretrained backbone. Custom architecture, data augmentation pipeline, checkpoint management, and training loop with learning-rate scheduling. Benchmarked against transfer-learning baselines.
Generative adversarial network for face synthesis — built from scratch in PyTorch. Custom discriminator/generator architecture, training stability techniques (label smoothing, gradient penalty), and mode collapse mitigation. Evaluated on FID score and visual coherence.
Production-grade evaluation harness for RAG systems and LLM pipelines. RAGAS metrics, deterministic route/source gates, per-question thresholds, sealed holdout sets, and impossible-score guards. Designed to prevent evaluator confidence from hiding product quality issues.
LangGraph agent that orchestrates specialized ML models and APIs as tools — decides which tool to invoke, interprets structured results, and synthesizes evidence into grounded responses. Demonstrates the "agent orchestrates specialized models" production pattern.
Deep dives into the architecture decisions, evaluation challenges, and engineering trade-offs behind these production systems.
Every layer used in these deployed systems — chosen for production, not notebooks.