Problem
Saroya AI, built at TinyCheque, lets creators import their content so their AI avatar can answer questions about it. Content arrives from X, Reddit, YouTube, LinkedIn and file uploads (PDF, DOCX, XLSX, PPTX). Every piece of creator content needed a verifiable hard-delete path across every data store.
What I did
The ingestion pipeline normalises each source into one document format, with OCR (optical character recognition) for image-heavy uploads. Every document is tracked through its lifecycle: ingestion, embedding, indexing and deletion. KYC identity checks verify a creator before their content goes live.
Hard-delete workflows run across Postgres, the pgvector index, object storage on Cloudflare R2 and the Redis cache, and are auditable.
Architecture
Limitations
The deletion audit is manual today. OCR quality varies with document format and image resolution. Source code is confidential.