I build AI systems that have to survive outside the notebook.

Problem

Technical information is distributed across different documents and needs fast, grounded retrieval.

My role

Owned delivery end to end across document ingestion, hybrid and multimodal retrieval, reranking, agentic retrieval, backend services, and frontend integration.

Outcome / learning

Fast mode achieved 5 seconds average latency in the internal demo.

Problem

Recognize small carton logos through existing CCTV in a production environment.

My role

Consulted on nearly 50 computer-vision use cases across factory sites and delivered this customer PoC with a two-person team, from requirements and data collection through model training, web app integration, Linux deployment, and on-site iteration.

Outcome / learning

With two cameras, a CPU-only server using a 12th-gen Intel Core i5 sustained 15 FPS. Under field conditions, accuracy was at least 83% and response time was under 150 ms, but the PoC did not meet its requirement of at least 95% accuracy.

Constraints

• Logos only a few dozen pixels wide
• Customer’s existing CCTV system
• Two cameras
• CPU-only server using a 12th-gen Intel Core i5
• Field issues involved camera quality, compression, perspective, and environmental conditions
• Bitrate capped at 1024 Kbps, which was low and often caused blurring
• Two Full HD cameras running simultaneously at a minimum of over 5 FPS

Problem

Connect face-recognition camera events with physical access-control workflows.

My role

Implemented backend APIs and fine-tuned custom object-detection models for additional customer use cases as part of a three-person team.

Outcome / learning

Contributed backend and model work to the production product.