Project 01 / 07 · Current role · since 2021
ViewMind
Building production software for cognitive assessment and AI-powered workflows.
- React
- TypeScript
- Python
- AWS
- OpenAI API
- MCP
- OpenClaw
- SQL
01 · Context
ViewMind builds software for cognitive assessment and neurological research. I joined in 2021 and I am the primary engineer responsible for application architecture and implementation. The work runs from the data and the models through to the applications people use, including CI/CD and production debugging.
02 · Problem
ViewMind supports cognitive assessment and neurological research, including memory, attention and executive function. Those applications have to stay correct once clinicians and research staff depend on them. Anything an AI workflow generates, or an internal tool can see, has to stay inside what that person is already allowed to access.
03 · My responsibilities
- Develop and maintain the production web applications, including the multilingual site at viewmind.com
- Set the application architecture
- Build data-processing systems
- Build production LLM workflows
- Design and deploy internal multi-agent tooling, including a custom MCP server
- Manage CI/CD and application deployments
- Troubleshoot production issues
- Guide engineering interns
04 · Architecture
A customer-facing workflow sits next to an internal multi-agent system. The workflow is a controlled call to the OpenAI API, with checks around it. The internal system is OpenClaw, connected to company tools through a custom MCP server.
Customer-facing workflow
- 01Structured dataApplication data comes in
- 02Controlled promptThe input is prepared before the call
- 03OpenAI APIThe model generates a response
- 04Deterministic checksFormat, structure, prohibited terms
- 05Automatic retryA failed check runs generation again
- 06Execution logSteps are stored for debugging
- 07Manual reviewA person checks meaning and content
Semantic review is not automated.
Internal multi-agent system
- 01Company toolsCRM, web analytics, support, internal APIs
- 02Custom MCP serverSelected API functionality only
- 03AllowlistsEndpoints, fields, and existing API roles
- 04Filter and aggregateReduce what can leave the API
- 05OpenClawThe system I deployed and connected
- 06External modelOnly the allowed slice is shared
I designed the deployment, integrations, server and access controls. I did not create OpenClaw or MCP.
Simplified public view. Endpoints, credentials and clinical data stay off this page.
05 · Technical decisions
- The customer workflow is not an autonomous agent. Deterministic checks cover formatting, structure and prohibited terms. A person reviews meaning and content quality.
- When a check fails, generation runs again. Each step is logged so the run can be debugged.
- The internal system runs on OpenClaw and reaches company tools through MCP. I designed the deployment, the integrations, the custom server and the access controls. I did not create OpenClaw or MCP.
- The MCP server exposes an allowlist of endpoints and fields, and it reuses the API role checks that already exist.
- Data is filtered and aggregated before anything is shared with an external model.
- The connected tools are the CRM, web analytics, customer support and selected internal APIs.
06 · Results
Primary engineer on the applications since 2021, including three production web applications used by clinicians and research staff. The LLM workflow and the OpenClaw system are deployed. I also guide engineering interns and handle production issues and releases.