AI Engineer building production Generative AI systems — agentic RAG, retrieval quality, and the LLM safety layers that make them shippable to enterprise customers.
Currently at AutomatR, working on an enterprise knowledge-base and search product deployed as multi-tenant SaaS and on customer-hosted infrastructure.
Agentic RAG Retrieval pipelines built as graphs rather than chains — LangGraph implementations of CRAG and Self-RAG, with query rewriting, decomposition, multi-source fan-out, relevance grading, hallucination checking, and an explicit refuse path when the evidence isn't there. I care a lot about systems that say "I don't know" correctly.
Retrieval quality Hybrid search over Elasticsearch — BM25 alongside dense kNN and multimodal embeddings, fused with Reciprocal Rank Fusion. Cross-encoder reranking, MMR diversity, calibrated relevance floors. Most RAG failures I've debugged were retrieval failures wearing a generation costume.
LLM safety & evaluation Guardrails mapped to the OWASP LLM Top 10: prompt-injection screening, PII/PHI redaction, output guards, and authorization checks that survive contact with a multi-tenant system. Paired with red-team evaluation corpora and CI gates — a guardrail you haven't measured on adversarial and benign inputs is a guess, not a control.
Document AI & computer vision Object detection for document structure (YOLO, model ensembling), OCR pipelines, and vision-language models for figure and table understanding.
Languages Python · TypeScript · SQL
GenAI LangGraph · LangChain · CrewAI · MCP · vLLM
Retrieval Elasticsearch · LanceDB · Pinecone · ChromaDB · FAISS
ML/CV PyTorch · YOLO · PaddleOCR · Hugging Face
Backend FastAPI · asyncio · RabbitMQ · Temporal · Redis · MySQL/MariaDB
Infra Docker · GitLab CI/CD · Azure · GCP · AWS
Most of the hard work in applied GenAI isn't prompting — it's retrieval correctness, evaluation you can trust, and failure modes that degrade honestly instead of confidently. The interesting bugs live in the seams between services, not inside the model.