TULON
A secure data access platform for machine learning researchers
Context
The product handled sensitive hospital and research data; I co-founded it and owned design and engineering.
When researchers work with sensitive data, there was no systematic way to govern who could reach what. And with medical data security is the gate rather than a feature — no hospital opens an adoption discussion before a product clears its review, so the bar had to be cleared before anything could be sold.
How I solved it
- An early-stage startup would normally start simple, but we were aiming at a major hospital for the first deployment, so the system had to hold to that standard from day one
- Shaped a domain-driven architecture and moved the platform to microservices accordingly
- Built the back office web service for data access governance
- Built server features: request and approval workflow, policy management, audit logging
- Built the isolated training environment
- automated VM provisioning and teardown, environment image templates, network isolation and firewall policy
- Built data egress control
- blocked exfiltration paths, a review flow for releasing results, session activity logging and monitoring
- Integrated hospital SSO and managed VM resource quotas
Result
- Built the platform that governs access to sensitive data, plus the isolated training environment it runs in
- request, approval, audit trail
- Two of Korea’s four major hospital groups were lined up as deployment sites as the build wrapped up