I build the parts around the model that make it useful: retrieval that can be checked, guardrails that decide what an agent may do, conversations that survive a hand-off between agents, and tool calling fast enough that people keep using it.
in production
Enterprise agents at SPS Commerce — RAG pipelines, guardrails, and multi-agent conversations.
multimodal retrieval
Product images and text embedded into one space, so retrieval stops caring which modality the answer lives in.
tool calling, tuned
Agents that act through MCP and tool calls — profiled and optimized, because a slow agent is an unused agent.
The day job: agentic systems for an enterprise commerce network — RAG pipelines, multimodal embeddings that put product images and text in the same index, and the guardrails and hand-offs that keep multi-agent conversations on task. Tool calling over MCP, kept fast.
Twitch chat is hostile input: strangers, mods, quotas, and a model that sometimes guesses. zenn_ai routes actions through permission gates, channel memory, local-model fallback, and an MCP sidecar.
A notes assistant that shows its sources as it answers. When retrieval looks weak, the graph reformulates the question and searches again instead of bluffing.
An Astro site with a streaming Cmd-K assistant exposed to the public internet, kept boring with per-IP quotas, Turnstile, KV caching, and OpenRouter.
model-training · token-counter · langchain-rag-app · agents · zengineer.cloud · exile-view
AI / ML
Languages
Web / Infra
Tell me what has to work, what can go wrong, and who has to trust it. I read every message.