Skip to content
Vesh AI

Private pilot · contract review

An AI that never lets the file leave.

It read 160 contracts in 128 minutes on a laptop. Once its models are installed, the review path calls nothing outside the machine — the material you are forbidden to send anywhere never has to go anywhere.

Request a briefingSee the measured evidence
The Vesh AI contract evidence matrix: one row per contract, with columns for parties, effective date, expiration date, governing law, assignment restriction and liability cap. Cells read Present, Not found or Uncertain, each with the quoted source text beneath. The status bar reads egress none.

The week you have now

Somebody still has to read every page.

The diligence set arrives on a Thursday. The tool that could have read it overnight is the tool your client's guidelines forbade, so it is read by hand — by the people you bill at the highest rate. What that costs is not the licence you did not buy. It is the weekend.

Look closer

What it does. And what it will not do.

  1. The first rows of the evidence matrix, with one cell outlined: the parties of the first contract, marked Present above the start of the source text it quotes.
    01

    Every answer shows its working.

    The cell opens the sentence it came from. A reviewer checks a citation instead of re-reading the contract, which is the difference between verifying work and doing it again.

  2. Four contract rows, with one outlined where three cells read Uncertain instead of asserting a value; cells in the row above read Not found.
    02

    It would rather say nothing than guess.

    An uncertain cell is marked uncertain and left empty. The afternoon goes on the cells it flagged, not on every page it read.

  3. The batch panel after a completed run of 72 cells across 12 documents, with the Export CSV, Export JSON and Destroy controls outlined together.
    03

    What comes into the room leaves with it.

    Export the matrix, then destroy the corpus. The text, the embeddings and the review decisions go; the record keeps file names and hashes rather than contents, and the destruction is written down.

Measured, not claimed

Not one invented citation.

Citations valid
100%
Positive claims resolving to indexed source text.
Precision
92.2%
When the system asserts a field, how often it is right.

960 review cells across 160 contracts it had never seen. Every claim it made resolved to real text in the real document. And when it asserts a field, it is right 92.2% of the time.

Internal held-out evaluation; not an independent audit. Measured on an Apple M4, 12 September 2026.

See how it was measured

Current status

A private-pilot MVP.

Stage
Private-pilot MVP for supervised contract review
Users
Single-user and single-tenant; no authentication yet
Deployment
Demonstrated on a single machine under Docker Compose
Benchmarked on
Apple M4 laptop with a 10-core GPU and 16 GB of unified memory
NVIDIA GPUs
Designed to run on NVIDIA GPU systems; NVIDIA GPU performance has not yet been benchmarked.
Not yet
Document-management integration, durable audit log, independent security audit
See the architecture

Private pilot

Bring contracts you are allowed to keep inside the room.

The first step is a bounded diagnostic: identify the review task, agree the held-out checks, and run the evidence matrix inside the boundary. Nothing is installed until you say so.

Request a product briefingSee the 14 days

Diagnostic brief

Day 14 of 14

  1. 01Where the material actually livesMapped
  2. 02What must remain a human signatureListed
  3. 03The runtime, and where it sitsNamed
  4. 04Risks we found, in plain languageWritten
  5. 05Price to install, price to keep runningQuoted

Yours to keep — for your board, your IT team or another vendor.