subtopic

Prove it handles the traffic

Write a Locust test that simulates realistic users (chat, uploads, predictions); run it against staging with 10 → 50 → 200 users; record latency p50/p95, errors and cost per minute; find the bottleneck (LLM rate limits, database, CPU, vector search) and fix it (more instances, connection pooling, caching, async). Set autoscaling limits from the results.