Projects
Things I’ve built, at work and on the side.
Side projects
Mercari AI Shopper
Five sequential agents turn a natural-language request into Mercari Japan searches, scrape listings with Playwright, and rank them with RAG. English and Japanese.
Research Copilot
A local RAG system with hybrid BM25 and dense retrieval.
LLM inference from scratch
GPT-2 inference built by hand: KV caching, prefill and decode, sampling and batching.
At work
Unified multi-LLM router
One service in front of Azure, OpenAI, Gemini and in-house models for HTTP and Kafka workloads. 1M+ requests a day at 99.9% uptime, and $100K+ a year saved through usage telemetry.
Agentic dashboard copilot
DAG-based orchestration across 200+ analytics sources, with an LLM-as-judge framework that caught a live regression after an upstream model update.
Parameter-aware rate limiter SDK
Fine-grained quotas across time buckets with zero-downtime config refresh, protecting LLM APIs from traffic spikes.
Publications
Evolution of Modern Deep Image Inpainting Methods
Co-author. A survey of deep learning approaches to image inpainting.