In development
In development · Early access

Never let a renewal slip through a PDF nobody re-read.

DeadlineGuard reads a folder of scanned agreements, extracts the dates that matter, scores how trustworthy each reading is, and chases the deadlines about to lapse. We're building it on our own document book first — and opening a small early-access list while we harden it.

deadline-guard
Read straight from the scan
Page-to-page cross-validation
Confidence scoring with named flags
Auditable, reproducible date logic
The problem

Operations teams sit on hundreds of dated documents — rent agreements, licences, certificates, compliance filings. The expiry date is buried in a scan, sometimes printed two different ways on two pages, sometimes mislabelled in the filename. Tracking them means a person opening each PDF, re-reading the fine print, and copying a date into a spreadsheet by hand. That doesn't scale, it drifts out of date, and one missed expiry can carry a real cost. The failure is almost never the deadline — it's that nobody re-read the document in time.

How it works

How it will work

01

Point it at your documents

It will pull scanned PDFs straight from a shared drive — no manual uploads, no renaming.

02

One AI read per document

A single multimodal model call transcribes the load-bearing fields — effective date, agreement period, document type, party names. No separate OCR step.

03

Deterministic date math + risk score

All expiry rules and arithmetic run in plain, auditable code — never in the model. Each document gets a 0–100 confidence score and named anomaly flags.

04

A ranked deadline queue

High-confidence expiries surface as “what's expiring, how soon.” Low-confidence reads are held back for a human instead of being guessed.

What's inside

Built for the real job, not a demo.

Read straight from the scan

PDF bytes go to the model directly as a document — no page-rasterisation, no brittle OCR preprocessing. (In build.)

Page-to-page cross-validation

When page 1 and page 2 disagree on a date, DeadlineGuard flags the disagreement for review rather than silently picking one. (In build.)

Confidence scoring with named flags

Every reading carries a 0–100 score and explicit anomaly tags — date mismatch, future-dated typo, suspiciously short duration, mislabelled file, corrupt scan. (In build.)

Auditable, reproducible date logic

The model only transcribes; all date math and renewal rules live in deterministic code, so the same document always produces the same computed expiry. (Proven on golden cases.)

Human-in-the-loop by default

Low-confidence documents never auto-write anywhere. Automated write-back stays disabled until a strict accuracy gate is cleared. (Locked design decision.)

Honest dev milestones

What's actually proven so far.

We're deliberate about what's proven versus planned. These are development milestones on our own data — not customer benchmarks.

  • Kerala expiry rule: proven 6/6 byte-for-byte against the legacy pipeline's golden cases — including the “X months and 29 days rounds up” registration quirk. A dev milestone on internal documents, not a customer benchmark.
  • One Gemini 2.5 Flash multimodal call replaced a two-stage Tesseract-OCR-plus-vision stack — a real pipeline simplification, validated internally.
  • Math runs in deterministic Python, asserted against golden tests — because the model isn't byte-deterministic even at temperature zero.
Who it's for

Made for teams who live in the work.

Operations teams with a big document book

Virtual-office and property managers, leasing operators, and anyone tracking rent-agreement renewals, licences or certificates at scale, where a missed expiry has a tangible cost.

The spreadsheet-babysitters

Teams who currently keep all of this in a spreadsheet someone has to manually keep alive — the exact job DeadlineGuard is being built to take over.

Under the hood

The stack

Gemini 2.5 FlashMultimodal document AIVertex AI (asia-south1)Deterministic Python rulesGoogle Drive ingestionQueue + worker batchConfidence scoring
Integrations
  • Document source. Designed to read scanned PDFs straight from a shared drive (Google Drive ingestion first). No manual uploads.
  • Output. A ranked deadline queue to a sheet or relational DB; automated write-back to master records is gated behind an accuracy bar.
FAQ

Straight answers.

Is this available yet?+

No — DeadlineGuard is in active development. The core engine and data layer run on our own server with all external writes off, and we're validating accuracy before any launch. Join the early-access list to help shape it.

Will it overwrite our records automatically?+

Not until it earns it. Automated write-back stays disabled behind a strict exact-match accuracy gate. Until then it runs in read-and-suggest mode, with a human in the loop for low-confidence documents.

Does the AI calculate the expiry dates?+

No. The AI only transcribes what's printed on the document. Every date calculation and renewal rule runs in auditable code, so results are reproducible and reviewable — not a black box.

What documents can it read?+

It's being built around scanned PDFs of dated agreements and compliance documents — rent agreements, licences, certificates. Early-access partners will help widen the supported types.

In development

Where this honestly stands

In development. Built and proven on Froshtek's own operations as customer zero. Phase 0 (the reading engine and data layer) is complete on our server with external writes off; the next phase — shadow mode against a live document source — is pending a service-account key. Automated write-back is gated behind a strict accuracy bar before any release. We're not selling it yet — we're inviting people who want it early.

Want DeadlineGuard early?

It's in active development on our own operations. Join the early list and help shape it — we'll scope a pilot with you, not sell you a license.

30 minutes · no obligation · we reply within 1 business day