Budget: 4000 USD Deadline: 14 days
Building a multi-tenant AI support platform where every conversation routes through an LLM, gets logged, and feeds back into quality metrics — the architecture decision that matters most here is whether the AI layer is synchronous (fine for low volume) or async with a task queue like Celery + Redis, which I'd go with to handle burst load without blocking FastAPI responses.
Concrete plan: FastAPI app with async endpoints, PostgreSQL for structured data (conversations, agents, tickets) plus pgvector for semantic search over your knowledge base. OpenAI API calls go through a worker queue with retry logic and idempotency keys so a failed webhook never double-charges or drops a message. Docker Compose locally, ECS or Fargate on AWS in prod, with secrets in Parameter Store.
One risk a junior misses: OpenAI rate limits under concurrent load. I'd add a token-bucket rate limiter per tenant and graceful fallback to a canned response so the platform stays responsive even when the API throttles.
What's the expected concurrent user volume, and do you need voice synthesis integrated from day one or is that a phase-two feature?
