Prismatic prepares a first-pass production set from a defendant’s document collection and the subpoena’s requests for production. A System 1 decision model scores every email, chat message, and attachment against the full wording of each request. The reviewer sets a cutoff and Prismatic tags the qualifying documents in bulk.
A Slack or Teams room can hold months of overlapping conversations in a short-message export such as RSMF. Prismatic scores each message with its surrounding conversation so it can interpret a reply like “Yes, that works.” Semantic search finds related messages posted days apart. I designed Prismatic around making a responsive message and the neighbors that explain it selectable as a single document for production.
In use
What makes it useful
194,688 scores for $3.90
Jev scored 16,224 synthetic records against 12 subpoena requests. The main batch covered 15,918 records in about 13½ minutes. The $3.90 covers model charges for scoring and evidence selection across the full corpus. Neither of the two deliberately unanswered requests received a tag at the demo cutoff.
Read the full document and override a tag
Open the full email and its attachments, the original file, or the chat in its conversation. Inspect the model’s score and supporting passage, then accept the tag, remove it, or add one by hand. Prismatic records each decision.
Catch tags outside the request’s scope
Prismatic checks tagged documents against the request’s dates and document types, including when a reviewer tags by hand. The scope queue explains why an automatic tag falls outside those limits and lets a reviewer keep or remove it.
Find the conversation inside a chat room
Each message has its own record with its author, timestamp, and reply links where available. A single search can surface a proposal and the reply that approved it days later. Each result opens in the original conversation.
See where the tone changes across teams
Compare each department’s and person’s usual share of negative messages with the weeks around a disputed event. A shift in support followed by changes in engineering, legal, or sales gives reviewers a place to look. Open the messages for a person and week to investigate what changed.
Under the surface
Open a layer. Explore the decisions inside.
One scoring call per document
Prismatic sends Jev the document, matter background, and all the responsiveness questions in one call through OpenRouter’s Decisions API. Jev evaluates the questions in parallel and returns a score for each request. A second pass can select a verbatim passage for a potential match.
Built with
TypeScript
React
Vite
MUI
NestJS
PostgreSQL
pgvector
Gemini Embedding 2
Jev
Supabase
Vercel
The throughline
I built Prismatic to replace e-discovery’s “find more like this” automation with a first pass guided by the subpoena itself. Each proposed document traces back to the request, the original record, and its tag history. Reviewers verify the final dates, document types, and content scope and complete privilege review before release.