An AI document-chat platform in production: upload documents, ask questions, get page-referenced answers.
- Role
- Lead Software Engineer
- Year
- 2025 — Present
- Focus
- AI platform · Backend architecture · Full-stack
Mindex IO
The problem
People work inside documents, not chat boxes. Mindex lets them upload files and talk to them, with answers that point back to the page they came from. The hard part is keeping that fast and reliable once real users and real files show up.
Role and scope
- Lead end-to-end architecture of a production AI platform
- Own an 85,000+ LOC TypeScript/TSX codebase across an agile team
- Containerized backend: Docker Compose, 6 services
- 295 API endpoints across 29 router modules, built for high availability
- Asynchronous task queues with Celery and Redis
- Dual-database layer (PostgreSQL and ChromaDB) for intensive AI workloads and real-time state sync
- Fault-tolerant third-party API integrations and multimodal ingestion, hardened for real-world deployment
Architecture
Key decisions
- 01
Keep slow work off the conversation
Interactive requests stay in the API. Heavy AI work runs on asynchronous task queues (Celery, Redis), so one slow document never blocks someone mid-chat.
- 02
Two databases, two jobs
PostgreSQL holds application state and ChromaDB serves vector search. The dual-database layer keeps intensive AI workloads and real-time state in sync.
- 03
Contain everything
Six services run under Docker Compose, serving 295 endpoints across 29 router modules, organised so each area of the API stays small enough to reason about.
- 04
Assume integrations fail
Third-party API integrations and multimodal ingestion are fault-tolerant and hardened for real-world deployment.
Outcomes
- A production AI platform with 295 API endpoints across 29 router modules
- An 85,000+ LOC TypeScript/TSX codebase owned end to end across an agile team
- A containerized backend of 6 services built for high availability
Stack
- TypeScript
- Docker Compose
- Celery
- Redis
- PostgreSQL
- ChromaDB