People rarely miss return windows, warranty coverage, renewals, or cancellation deadlines because they ignore their calendars. They miss them because the important date is buried in a receipt, invoice, warranty card, or subscription notice.
Deadline Desk was inspired by that frustrating gap: purchase documents contain valuable deadlines, but they are scattered, easy to forget, and often written in inconsistent language. We built a focused tool that turns those documents into deadline records users can actually act on.
A user uploads a document or selects a sample. AI extracts the purchase details, possible deadlines, confidence levels, supporting evidence, warnings, and missing fields. Instead of pretending the model is always correct, Deadline Desk asks the user to review and edit the result before saving it. The verified deadline then appears in an urgency-based dashboard, with reminders and calendar export for known dates.
The biggest lesson was that document extraction should not be treated as a source of truth. A receipt may state a purchase date but omit the retailer’s return policy; inventing a deadline would be worse than returning no answer. We designed the app to preserve uncertainty, show evidence, validate structured output with Zod, and keep unknown deadlines explicitly marked as unknown.
We built Deadline Desk with Next.js, TypeScript, Tailwind CSS, and a provider-isolated AI extraction layer using an OpenAI-compatible API. Codex and GPT-5.6 supported the implementation process: architecture, schema and prompt design, validation strategy, debugging, and submission preparation. The runtime extractor uses Cerebras Gemma 4 31B, while the architecture leaves room for an optional GPT verification pass.
The main challenges were handling inconsistent document formats, normalizing dates safely, and creating a workflow that is useful without overstating AI accuracy. The resulting product is a human-in-the-loop deadline assistant: AI does the tedious reading, while people keep control of the final decision.
Built With
- cerebras
- codex
- gemma-4-31b
- next.js
- railway
- tailwindcss
- typescript
Log in or sign up for Devpost to join the conversation.