I turn messy, real-world data into production AI.

Built LLM-powered systems for medical document parsing, structured extraction, and automated patient health summaries — N1's Comprehensive Healthcare Reports.

• Built a medical document router that cut downstream parser LLM calls by ~75% with no loss in extraction accuracy.
• Raised extraction quality through evaluation pipelines — deduplicating and normalising biomarkers across labs (so the same test from two providers maps to one clean record), adding medication coverage, and engineering prompts and context.
• Designed a LangGraph multi-agent system that generates these patient summaries from raw records — coordinating specialist agents for clinical analysis, literature research, narrative generation, and fact validation.
• Benchmarked 10+ LLMs, VLMs, and parsing tools (GPT, Claude, Gemini, Qwen, MinerU, Docling, PaddleOCR-VL) — choosing the right model per task on measured accuracy and cost, not assumptions.
• Ran prompt iteration, observability, and model comparison through Langfuse and LangSmith.

Computer vision and operations AI for a convenience-retail chain — from edge cameras to live Power BI dashboards.

• Architected and deployed a Ready-To-Eat product-availability system across 2 pilot stores, with a scalable rollout underway.
• Containerised inference on Kubernetes-on-AWS-EC2 with a central GPU cluster — avoiding ≈RM 5,000 of hardware per new store.
• Built an end-to-end data pipeline (edge → S3 → Python ETL → Power BI) giving managers live replenishment KPIs.
• Shipped on-shelf availability detection, customer & product flow counting (multi-object tracking with re-identification to avoid double-counts), a food-fraud prototype, and cash-transaction recognition.
• Prototyped an autonomous shelf-auditing robot — ROS2 navigation with SLAM map-building for aisle patrol — plus an internal policy Q&A assistant using retrieval-augmented generation (RAG).