Tessera
Document intelligence that reads contracts, extracts what matters and cites every answer.
This is a concept exploration created by Senume. It does not describe a client engagement, and no results are claimed as real.
The problem
Review teams spend hours locating the same clauses across hundreds of documents. Generic AI chat tools help, but without citations their answers cannot be relied on for professional work.
- Role
- AI architecture and product build
- Timeline
- 10 weeks (concept)
- Services
- Applied AI, Engineering, Experience Design
- Platforms
- Web
- Technology
Constraints
-
Every generated statement must be traceable to an exact passage.
-
Client documents cannot be used to train third-party models.
-
Reviewers need to correct the system and see the correction stick.
Solution
Tessera is built around a simple rule: no answer without a source. Documents are split into semantically meaningful passages, embedded and indexed. Each question retrieves the most relevant passages, and the model is constrained to answer only from them — with inline citations the interface turns into highlights.
The review workspace treats AI output as a draft. Reviewers accept, edit or reject extractions, and those decisions become an evaluation set that keeps quality measurable over time.
How it flows
- Documents
- Chunk & embed
- Retrieve
- Generate with citations
- Review
Results
Intended outcomes — concept project
-
Cited by default
Each answer highlights the source passage so reviewers verify in one click.
-
Structured extraction
Key terms land in a reviewable table rather than a wall of chat text.
-
Human in the loop
Corrections feed a review queue that improves future extractions.
What we learned
- Trust in AI products is earned through interface design as much as model quality.
- Retrieval quality mattered more than model size for this domain.