Urgent.News

What's breaking now, across thousands of outlets.

Tech

Shoebox: a small model reads your receipts overnight, slowly, and never guesses a number

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 3: Slow and Steady What I Built Receipts are the paperwork nobody does on time. They go into a wallet, then a drawer, then a shoebox, and at the end of the month (or the end of the tax year) someone sits down and types them into a spreadsheet one by one: date, shop, total, category, then files the paper or the photo…

This is an open-source AI challenge submission for the third week called Slow and Steady. Shoebox is a folder that organizes your receipts while you sleep. You just take photos of receipts whenever you get them and drop them into Shoebox's inbox. At night, Shoebox processes the new photos, one at a time, on your own computer using the Gemma 4 model.

The model examines each photo and copies down every line printed on the receipt, including the total, discount, tax, and change. It confirms the total by checking other figures on the same receipt. If the total checks out, the receipt is filed with a name that includes the date, shop name, and amount. If the total doesn't match, it's flagged and left in a "Check me" folder for further review.

Shoebox only adds a receipt to your expense sheet once it's been confirmed by the model. If a receipt can't be processed automatically, it waits until the morning and gets a note indicating that it needs to be checked. Shoebox runs locally on your computer, so you don't have to worry about anything. After processing, Shoebox provides a summary of what it did overnight, including the total amount spent and any receipts that still need your attention.

Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.

Read the original at dev.to →

More in Tech

Your team has answered this before. I’m building a tool that remembers.

I built a tool to help with community support. Then I kept building. I spend a good part of my day working with the community at CopilotKit, answering questions in Discord, GitHub, and Slack, helping…

  • answerLoops consolidates community support questions from GitHub, Discord, and Slack.
  • Knowledge base populated via direct entry, documentation import, file upload, or syncing files.
  • Resolved tickets can be turned into knowledge-base articles with Promote to KB action.

CodeDrill: Master Java Collections and DSA Through Deliberate Practice 🧠💻

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 3: Slow and Steady . What I Built Meet CodeDrill — a deep learning and practice platform designed to help developers go beyond…

  • CodeDrill platform enhances Java Collections and DSA understanding through deliberate practice.
  • Focuses on Java Collections internals and structured DSA practice with spaced repetition.
  • Open architecture and learning resources designed for community contributions and adaptability.

Tokenization is why your agent bill looks weird

Models don't read text. They read integers. Between your string and the model there's a tokenizer, and it quietly decides how much you pay, how much fits in the context window, and why your model is…

  • Tokenization divides strings into vocabulary pieces from 30k-200k entries
  • Byte Pair Encoding (BPE) merges frequent character pairs into new symbols
  • Hindi tokenization differs significantly from other languages, using 24 tokens vs 111

The KV cache is eating your GPU. MQA and GQA are how models fight back

Quick recap from the last post: the KV cache saves you from recomputing old tokens, and it's the reason generation is fast. But it has a bill. And the bill is memory.

  • KV cache consumes 512 KB per token for Llama-2 7B
  • MQA reduces memory usage by 32x compared to traditional MHA
  • GQA balances MQA and MHA, reducing cache size by 4x

More from Tuesday 20 October →