Architecture (reference summary)
High-level architecture infographic (same asset as the Technology page).
Canonical vs. derived state
PostgreSQL remains the system of record for users, permissions, billing, final scores, and checkpoints. Amazon OpenSearch holds derived full-text, vector (kNN), and analytics views. If search drifts, rebuild the index from Postgres—truth does not live in the cluster.
Managed services
The reference stack pairs AWS Amplify for edge delivery, S3 with presigned uploads, Cognito for identity (including Marketplace-native paths), Textract for OCR, and Bedrock (Converse API) for structured LLM scoring.
Pipeline shape
Intake creates a queued row, workers acquire bytes and hash them, Textract extracts text, enrichment adds metadata, Bedrock produces structured scores, results persist to Postgres, OpenSearch indexes asynchronously, and the UI receives completion via SSE.
Failure handling
Malformed JSON is incrementally parsed and validated; search outages defer indexing; worker crashes resume from checkpoints; Redis loss falls back to polling with Postgres-backed event logs.