Project 01 / 07 · Current role · since 2021
ViewMind
Leads three production web apps for clinicians and research staff, and the research data platform underneath them.
01 · Problem
ViewMind measures memory, attention and executive function from eye movements, for neurology and pharmaceutical studies. In that setting a wrong number in a clinical report is not a cosmetic bug.
02 · Approach
Designed the research data platform for data quality, query performance and a schema that held together years into a longitudinal study. Shipped on AWS Lambda and S3. Also built the multilingual marketing site at viewmind.com.
03 · Outcome
Three production web applications led, used by clinicians and research staff.
Recent work
Ops MCP server
Narrow, audited tools over the Model Context Protocol that let staff ask an agent operational questions. Every call runs with the asking person's own permissions and leaves an audit line with IDs only. Clinical data, scores and reports never become tools, and anything that writes waits for a person to approve it.
Agent harness
Changes to the agent's instructions or tool descriptions are tested before they reach production. A probe replays a fixed set of real staff questions through both versions and records which tool was called, whether the agent asked a follow-up when it should, and whether its answer repeated the tool's number exactly. The better version ships with that evidence.
Related note: A green trace is not a correct outcome →Customer-facing AI agent
An assistant inside the ViewMind portal that answers customers' operational questions, such as completed evaluations, report readiness and usage. It reads through the same session the person is signed in with, so it only sees what they can already see.