Role focus / AI product engineering

AI Product Engineer

I build retrieval-grounded products with bounded agents and human approval at the consequential steps.

OptiTalent: independent project, no users yet
Muriersilk concierge: client delivery, production deployed

01 / MCP

The assistant cannot approve itself.

OptiTalent exposes product capabilities through Hosted MCP. Sensitive actions stop for exact-input browser approval.

Hosted MCP validates assistant requests; sensitive actions require an exact-input ticket and separate browser approval — Implemented access and approval boundary. OptiTalent has no users yet. Open the diagram for full size.

Read the OptiTalent case →

02 / RAG + agent workflow

Retrieval first. Model authority second.

The Muriersilk concierge combines dense and keyword retrieval, bounds the candidate set, then lets the model select stable product IDs. The browser loads current commercial facts; a person can take over.

Muriersilk concierge showing real silk glove product cards after interpreting an imperfect customer request — Production customer surface. The screenshot demonstrates a bounded product recommendation, not a traffic or revenue result.

Read the retrieval and agent case →

Browse all work →