Literay

Turns dense documents into a dialogue you actually understand.

Comment

Inspiration

Every time we sign things we don't fully understand — lease agreements, terms of service, internship contracts, grant conditions. The clauses that matter most (liability, termination, indemnification) are usually the ones written to be skimmed past. Existing AI chatbots can summarize a document in one shot, but a wall-of-text summary doesn't actually teach you anything, and the moment you close the tab, all context is gone. The next contract, you start from zero again.

The Collaborative Partner track description crystallized what we'd been circling around: an agent that doesn't just answer once, but leads the way and takes notes — asking clarifying questions, guiding step by step, and remembering how you personally struggle so it gets better at explaining things to you specifically. That's the gap we wanted to fill: not a smarter summarizer, but a reading partner with a memory.

What it does

Literay walks a user through a dense document clause by clause instead of dumping a single summary. For each clause it:

  1. Explains the clause in plain language, grounded in the actual source text (RAG over the uploaded document, so it can't drift from what's really written)
  2. Asks a clarifying question when the clause depends on the user's specific situation ("is this a residential or commercial lease?")
  3. Checks understanding with a short comprehension question before moving on
  4. Logs which concepts the user struggled with — not just this session, but across every document they've ever brought to Literay

The next time the user opens a new document, Literay checks its memory first. If it knows they've repeatedly struggled with indemnification clauses, it slows down and over-explains that clause before being asked — the same way a human tutor would recognize a recurring weak spot.

How we built it

  • Gemini as the reasoning core for clause explanation, clarifying-question generation, and quiz generation
  • Agent Development Kit (ADK) to orchestrate the multi-step loop: ingest → explain → clarify → quiz → update memory, rather than a single prompt-response turn
  • Vertex AI Search / RAG grounding over the uploaded document so every explanation cites the actual clause text instead of hallucinating contract language
  • Agent Engine Memory Bank for the persistent, cross-session memory layer — this is what lets Literay recall a user's weak spots in a document they open weeks later, not just earlier in the same chat
  • Cloud Run for deployment, so the agent scales to zero between sessions and doesn't cost anything when idle

What's next for Literay

  • Support for comparing two related documents at once (e.g. "how does this lease differ from your last one?")
  • A confidence-calibrated flagging system that surfaces clauses that are unusually one-sided compared to standard contract language

Built With

Share this project:

Updates

Submission history