Stanford Securities Litigation Analytics
An AI model reads a securities litigation filing and fills the schema for its document type, citing verbatim evidence per field. A human reviewer verifies every field against the source PDF before it is accepted. Nothing ships unread.
How it works
Drop in a filing PDF and pick its type. Scanned-doc and draft detection up front.
The model fills that document type's schema with verbatim evidence for each field.
A guided viewer jumps to each highlighted passage. Approve field by field.
Entity-normalized, validated, and queued to the downstream pipeline.
Features
Every field carries verbatim evidence; a coverage threshold flags low-grounding extractions before review.
The viewer steps between highlighted source passages one at a time, so each value is checked against the page it came from.
Nothing is auto-accepted. A per-field approval flow keeps a human in the loop on every value.
Issuers and parties are fuzzy-matched against a live catalog, with same-name and stale-cache guards before submission.
Work auto-saves; a draft picker lets reviewers resume mid-review across sessions and devices.
Likely-scanned filings with no extractable text are rejected up front instead of producing empty extractions.
A registry maps each document type to its extractor and schema. New types slot in without touching the core flow.
SAML single sign-on with an individually managed reviewer allowlist.
Supported document types
Architecture
A Flask backend coordinates the pipeline; a Vue review surface runs the per-field approval flow over a PDF.js viewer. Extraction calls the model with a per-document-type system prompt and a JSON-prefilled response, then maps evidence spans to passage coordinates for the viewer.
Storage, the submission ledger, the entity cache, and secrets sit behind a backend-adapter seam: a filesystem/JSONL/in-memory implementation for local development and an AWS implementation (S3, DynamoDB, Secrets Manager, CloudWatch) for deployment, selected by one environment switch. The same application code runs in both.
It is deployed as a container on AWS Fargate behind an Application Load Balancer with HTTPS, authenticated by Stanford SAML SSO, with structured telemetry and per-extraction cost tracking throughout.
Get involved
Reviewing means reading a securities litigation filing alongside the extracted fields, then confirming or correcting each one against the source. The work is close to a paralegal's: careful reading of complaints, attention to the theory of liability, and a working sense of how securities cases are built.
People who find this interesting often have a footing between undergrad and law school, or are otherwise looking to build hands-on experience in securities law. The project is part of Securities Litigation Analytics at Stanford Law School, led by Professor Michael Klausner, a leading authority on securities litigation. We are happy to tell you more about it.