Stacks/AI product infrastructure

An anonymous startup's production AI infrastructure (HN)

On an Ask HN thread, k9294 shares a full production stack from a small startup: an OpenAI-compatible gateway with fallbacks and cost tracking, a homegrown job queue they already want to replace, OTEL plus SigNoz (no LLM-output logging), Postgres and S3 for memory, LlamaIndex and Qdrant for RAG, ClickHouse via Tinybird for usage analytics, and Stripe meters for billing. The company name is withheld.

k9294Mar 16, 2025Original writeup ↗
Use case
Production AI infrastructure at a relatively small startup with a few thousand monthly active users, covering chat, transcription, image generation, embeddings, RAG, and user-facing tool integrations
Category
AI product infrastructure
Team size
A relatively small startup (headcount not stated); a few thousand monthly active users
Orchestrator / framework
The team started with LangChain and moved off it. They now use the OpenAI SDK plus a custom API gateway that handles cost tracking, metered billing, model fallbacks, aliases, and feature flags (structured output, reasoning, modality). Workflow orchestration is a simple task executor and background job queue for retries and durable execution; the author says this has become a limitation and they are considering Temporal, trigger.dev, and Conductor, with no chosen winner yet.
Memory
Postgres for agent memory, chat history, and persistence (a chat-app-style schema of threads, messages, and attachments), plus S3 for files and images. RAG is a custom LlamaIndex ingestion service with managed Qdrant for vector search.
Evals
not stated
Observability
OpenTelemetry and SigNoz. The team does not track LLM outputs, for security reasons, because the product is a consumer app with a lot of private data.
Monthly cost
not stated
Observability tools
Other tools
ClickHouse via Tinybird (usage and cost analytics)Stripe meters (evaluating Orb as a replacement)LlamaIndexQdrantLLM tool wrappers and OAuth/API-key credential management