One AI product had to handle 18 distinct task types, and each pipeline needed validation and had to stay stable and observable. I owned the architecture, backend, frontend, AI pipelines, data layer and deployment. Claude and GPT pipelines run in parallel, with a queue, retries and failure handling, structured JSON output and PDF reports. The platform went from its first release to 50+ companies with under 3% churn.