AI systems engineering: orchestration, local inference, and tool routing built to survive contact with production.

Local-first LLM orchestration with a cloud fallback

Qwen 2.5 14B on consumer ROCm hardware at ~63 tok/s, Claude Haiku as fallback, a modular tool registry, and an intent routing table that separates direct from contextual intents.

• Qwen 2.5 14B
• ROCm
• Claude Haiku
• tool registry

A companion that actually knows your library

A streaming chat that tool-calls real catalogs for titles you don't own yet, keeps a persistent taste profile, bridges anime, manga and books, and returns actionable cards — Claude Haiku by default, Sonnet on demand.

• Claude API
• NDJSON streaming
• taste memory

An in-product AI assistant and content ranking

Designing the in-product AI assistant and the content-ranking logic that shapes what creators and audiences see.

• LLM orchestration
• content ranking

Claude inside a real product: assistant, vision, pricing

A Claude-powered assistant answers project questions, vision receipt scanning turns a photographed receipt into line items, and price suggestions help fill the estimate — practical LLM integration wired into the domain model, not bolted on.

• Claude API
• vision receipt scan
• price suggestions

Agent workflows that hold up across six projects

A queue-table pattern — a deliberate schema constraint — removed a whole class of agent failures in property filtering, making the pipeline reliable enough to run daily across six active projects.

• MCP
• Claude Code
• queue-table pattern