Deadline Desk

Inspiration

A deadline alone does not finish a return. The receipt, policy, next step and follow-up often live in different places. I designed a personal workspace that connects those steps while keeping uncertain evidence visible. This is a product hypothesis; I have not yet measured real-user impact.

What it does

Deadline Desk accepts receipt and policy text. A local model suggests fields with source quotes, and the person reviews them before saving. Missing delivery evidence and conflicting windows remain unresolved. A voice or text request can prioritize purchases, explain a deadline, prepare a checklist, draft a merchant message, export a calendar file or start a confirmation to record a completed return. Records persist in the browser and can be reopened.

How I built it

I built a Node.js web application with browser storage, Qwen3-4B through llama.cpp and Whisper base.en through whisper.cpp. The language model classifies requests into seven bounded intents and suggests receipt fields; deterministic application tools calculate dates and produce actions. Microphone audio is transcribed locally and temporary server recordings are deleted after processing. A transcript is reviewed before the request is sent. Read-aloud uses the browser's speech synthesis.

This is an independent Alexa+ web simulation with its own working voice interface, not an Amazon-hosted integration. I did not use the gated Alexa+ SDK or claim native Alexa access. The MIT source and separate model download instructions make the prototype locally runnable.

What I validated

I verified receipt review, priority selection, checklists, confirmation, persistence after reload and reopening in the actual browser. Eight boundary checks pass. A limited held-out text intent set passed 20 of 20 examples. Six clean synthetic speech recordings passed transcription-to-intent checks. A further 12 StepFun-generated clips with emotion, accent and inline delivery requests reached the expected intents; 9 transcripts matched the requested text after normalization. Two transcripts repeated a sentence and one omitted a filler. These are synthetic pipeline checks with zero human participants; accent fidelity and the source of repetitions have not been isolated. Silence is rejected before transcription; oversized audio and requests from a foreign browser origin are rejected. These results are limited checks, not general accuracy, accessibility or user-impact claims.

Challenges and next steps

An initial smaller model misclassified some intents, so I changed to a larger local model. Speech recognition produced text for silence during a failure check; an audio-energy gate now rejects silent or very quiet input. Quote validation alone does not understand every policy exclusion, so review remains explicit. Human microphone quality, image/PDF extraction and real-user impact still need validation. Calendar files and drafts are exports, not connected account actions.

Source

https://github.com/widechaos/deadline-desk-alexa-demo

Built With

• ffmpeg
• javascript
• kokoro
• node.js
• qwen
• stepfun
• whisper