Building production AI systems that actually ship.

Production RAG system for real-time tennis analytics.

• Production RAG system for real-time tennis analytics
• Built 5-stage pipeline with query routing, hybrid retrieval, and streaming generation
• Reduced LLM latency from 3s to 400ms (7.5x faster) using chunked streaming
• Integrated via FastAPI endpoints.

Multi-agent AI system for autonomous code generation.

• Multi-agent AI system for autonomous code generation
• Four specialized agents (Planner, Coder, Reviewer, Tester) orchestrate complex coding tasks with sandboxed E2B execution—reducing iteration time by 60% vs manual development cycles
• LangGraph provides state management for reliable agent coordination, solving the coordination failure problem (30% → <5%).

Production MLOps platform for real-time fraud detection.

• Production MLOps platform for real-time fraud detection
• Sub-100ms inference (5x faster than industry standard 500ms) with automated retraining on drift
• Handles 10K+ predictions/sec with Grafana observability
• Why it matters for 2026: Same MLOps patterns power LLM guardrails—fraud detection principles transfer directly to prompt injection detection and output validation for GenAI systems.

Intelligent LLM router with semantic caching.

• Intelligent LLM router with semantic caching
• DistilBERT classifier achieves 97% routing accuracy at 50ms—routes 70% of queries to GPT-3.5, 30% to GPT-4 based on complexity
• Result: 60% cost reduction vs 100% GPT-4 baseline
• Hybrid cache strategy (exact-match Redis + semantic Qdrant) dropped false positives from 15% → <2% through threshold tuning.

Industrial RAG with hybrid retrieval (BM25 + dense embeddings) and cross-encoder re-ranking for enterprise document search.

• Industrial RAG with hybrid retrieval (BM25 + dense embeddings) and cross-encoder re-ranking for enterprise document search
• Hybrid approach: 94% relevance vs 78% dense-only baseline (+16% improvement)
• Architecture choice: BM25 catches exact keywords, SBERT handles semantic matches—cross-encoder re-ranking on top-50 adds final 4% boost
• Evaluated on 2K query-document pairs with manual relevance judgments
• (Private enterprise deployment)