Fifty users to a thousand.
Led 5 engineers rebuilding a research LLM platform into Aideator, a production system.
Senior R&D Engineer · Technical Lead of a 5-person cross-functional team · Oct 2025 – present
Context
Deliberatorium is a UM6P research platform that uses LLMs to structure large-scale deliberation. It existed as research code. I lead its rebuild into Aideator, a production system.
Problem
The legacy system topped out around 50 concurrent users, its research concepts weren't productised, and years of historical deliberations had to survive the move.
Architecture
- Participants
- React · TypeScript client
- FastAPI · LLM features
- PostgreSQL · Redis
- Docker Compose on DigitalOcean
Key decisions & trade-offs
Re-architect, don't patch
Rebuilt the full stack on React/TypeScript, FastAPI, PostgreSQL and Redis.
Trade-offMore work up front than incremental fixes, in exchange for removing the scaling ceiling.
Migrate the history
Owned the end-to-end migration, ingesting historical data while re-implementing research-paper concepts as production LLM features.
Trade-offA harder migration than starting fresh, but research continuity is the platform's value.
Standards before features
Set Docker-based deploys, code review and CI practices, adopted as the default across the project.
Trade-offSlower first weeks; every engineer after that ships the same way.
Outcomes
- ~50 → 1,000+ concurrent users (20×)
- High uptime with real-time incident detection
- Historical data preserved through the migration
- Team standards adopted project-wide