The service foundations
I designed asynchronous OCR and transcription services using FastAPI, Kafka, and AWS. Provider integrations live behind internal interfaces. Job claims, retries, bounded concurrency, and result delivery are part of the service contracts rather than afterthoughts.
A decision that mattered
In the summarization MVP, provenance is attached before an LLM sees the text. The model refers to evidence IDs; the application resolves the document and page. Citation excerpts are then checked against their referenced evidence. Missing information remains visible, and human review is required.
Measuring the claim
An early evaluation on four public appellate decisions, totaling 45 native-digital pages, measured extraction citation rejection at 3/177 (1.7%) and narrative citation rejection at 2/54 (3.7%). Both passed the project’s pre-set 5% threshold. These are citation checks on a small corpus, not a measure of overall legal accuracy or production performance.
Where the work stands
The summarization service is an MVP, with rollout work still remaining. Its durable workflow uses LangGraph and PostgreSQL checkpoints, with per-job cost controls. Noisy OCR robustness and broader legal validation need further evidence before wider use.
I want a reviewer to see what the system knows, where it found it, and what it could not establish.