DEV Community: Maltheshhh

Slow Digest: a local open-weight model reads your group chat overnight and leaves you two lines

Maltheshhh — Sat, 24 Oct 2026 17:48:58 +0000

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 3: Slow and Steady

What I Built

Slow Digest turns a long group chat into two lines. Paste or export the chat, and it returns the most important change (what, when) and the most important thing you must do (who, by when). The sticker packs, match talk and goodnights are skipped.

It is for anyone who wakes up to 200 unread messages in a class, club or project group and only needs to know what is due. You can run it from the web page, or schedule a script that writes the digest to a file while you sleep.

Demo

Live site: https://slow-digest.onrender.com

The hosted site runs on Render's free plan, which has no Ollama. It uses the rule-based digest there. The open-weight model runs when you start the app on your own machine.

Code

maltheshhh / slow-digest

a “Sloth Summarizer” that squeezes long chats into two lines using an open-weight model that runs in your browser.

Slow Digest

A two-line morning summary of your group chat, written by an open-weight model running on your own machine.

Built for the DEV Hacktoberfest Open-Source AI Challenge, Week 3: Slow and Steady.

Paste an exported chat, get two lines: what changed, and what you must do. Jokes, stickers and goodnights are skipped.

Why open

Chats stay on your machine. The server only talks to Ollama on localhost.
Works offline. If no model is installed, a rule-based digest takes over.
Swap the model. Set OLLAMA_MODEL or choose one per request.
No cost per run. No API key.

Stack

Node 22 + Express (API and static hosting)
Ollama (any open-weight model it can serve)
SQLite via the built-in node:sqlite for digest history
Plain HTML, CSS and JS frontend

Run it

ollama pull llama3.2:3b     # optional, enables the model engine
npm install
npm start                   # http://localhost:3000

Copy .env.example to .env and export…

View on GitHub

How I Built It

Model: any open-weight model that Ollama can serve. The default is llama3.2:3b, and you can change it with one environment variable or per request.
Backend: Node 22 and Express. POST /api/digest sends the chat to Ollama on localhost and asks for exactly two lines. If Ollama is not running, the same endpoint falls back to a rule-based digest that scores messages by deadlines, days, times and numbers.
History: stored in SQLite through Node's built-in node:sqlite, so there are no native packages to install.
Frontend: plain HTML, CSS and JS. A tiny sloth crawls along a branch while you read, at an assumed 4 metres a minute.
Overnight mode: scripts/sloth_digest.py does the same job from the command line with the Python standard library only. Cron runs it at 3 am and the digest is in a text file by morning.
Tests: 5 tests cover chat parsing, the digest rules and the API.

Why Does Open Innovation Matter?

A group chat holds other people's messages, so I did not want to send it to a server I do not control. With a local open-weight model, the chat never leaves the laptop.

The same choice gave me three other things a closed API would not:

It works offline. The overnight job needs only a laptop, and the app still works with no model installed.
I can swap the model. A smaller one is faster, a larger one handles nuance better, and I can retune the prompt to how my friends write.
It costs nothing per run. There is no API key and no meter running while I sleep.

I built a receipt sorter that never uploads a receipt

Maltheshhh — Sat, 17 Oct 2026 17:23:19 +0000

Every freelancer I know has the same monthly chore: a folder of screenshots, PDFs and receipt photos that gets sorted by hand before the accountant asks. I built receiptsort to do that sorting, and I wanted it to run with no cloud at all.

Live site: https://receiptsort.onrender.com/
Source: https://github.com/maltheshhh/receiptsort

What it does

Point it at a folder of receipts and invoices. A vision model extracts vendor, date, invoice number, currency, subtotal, tax, total and category. Each file is copied to YYYY-MM/Category/ with a searchable name like 2026-09-03_Figma_12.00.png, and ledger.csv gets one row per receipt.

Anything uncertain goes to _needs_review/ instead of being filed with a guess:

a missing vendor, date or total
subtotal + tax that doesn't match the total
the same vendor, date and total as an earlier receipt

Files are tracked by SHA-256, so re-running on the same folder only processes new ones.

Why open models

Privacy. Invoices carry client names, rates and bank details, and many freelancers are under NDAs. The model runs locally through Ollama (default qwen2.5vl:7b), so nothing is uploaded and it works offline. A hosted API would exclude the exact users this is for.

Cost. Hosted vision APIs bill per page. Local inference is free per receipt, so a flat price for the software still makes sense in a busy month.

Swappable model. Any Ollama vision model works with --model. I can try a newer one the week it ships or fine-tune for a niche, with no vendor in the way.

The interesting part: schema-constrained output

Ollama accepts a JSON schema in the format field, which constrains the model's reply. That makes a 7B model dependable for extraction:

body = {
    "model": model,
    "stream": False,
    "format": SCHEMA,              # JSON schema with a category enum
    "options": {"temperature": 0},
    "messages": [{"role": "user", "content": PROMPT, "images": [image_b64]}],
}

The schema limits category to a fixed list, and a validation step checks dates, parses numbers and cross-checks the totals. Whatever fails validation is routed to review.

Testing without a model

The test starts a stub Ollama server on localhost that returns canned answers, then runs the whole pipeline: filing, flags, ledger and the idempotent re-run. It caught a real bug where duplicates inside a single run weren't flagged, because in-memory rows held floats while rows loaded from the CSV held strings.

I haven't yet measured accuracy on a large set of real receipts. That's the next step.

Try it

ollama pull qwen2.5vl:7b
pip install pymupdf   # optional, for PDFs
python3 receiptsort.py ~/Downloads/receipts --out ~/receipts-sorted

The website is a static page on Render. The sorter itself runs on your own machine, which is the point.

Feedback welcome, especially on receipts from other countries and languages.

Touch Grass: An App That Pays You to Go Outside

Maltheshhh — Sun, 11 Oct 2026 00:22:04 +0000

Why I built it
"Touch grass" is what the internet tells people who spend too much time online. I turned the joke into a game: finish small outdoor quests, prove it with a photo, and earn rewards.
Try it live at https://touchdegrass.onrender.com The source is on GitHub https://github.com/maltheshhh/TouchdeGrass.
How it plays
A round takes four steps:

Pick a quest, such as shoes off, tree hug or golden hour.
Upload a photo as proof.
Claude judges the photo and rejects screenshots, photos of screens and indoor shots.
A pass earns blades (the points) and a "Verified" stamp. Enough blades unlock vouchers. Features Touch Grass ships eight quests worth 155 blades in total, plus a progress meter and three vouchers. • AI photo judge: Claude checks each photo against the quest and leaves a one-line note on what it sees. • Photo proof: upload from your camera or gallery. Screenshots and photos of screens fail. • Blades and vouchers: a meter tracks your points, and vouchers unlock at 30, 80 and 140 blades. • Voucher codes: claim a code and copy it in one tap. • Saved progress: completed quests and claimed vouchers are kept in your browser. • Grass you can brush: the hero lawn bends away from your cursor or finger. • Light and dark themes: the lawn turns to night when your device is in dark mode. • Phone friendly: built to work at phone width, with a Start over button for a fresh run. The eight quests: Quest What to photograph Blades Level Shoes off Your bare feet on real grass 20 Easy Palm down One hand flat on a lawn 10 Easy Tree hug You hugging a tree, or both hands on the trunk 20 Medium Look up Open sky, with no ceiling or screen in frame 10 Easy Meet a local A bird, bug, snail or squirrel 25 Medium Leaf trio Three different leaves side by side 20 Medium Golden edge A sunrise or sunset 30 Hard Outdoor break A snack or a book outside 20 Easy Under the hood The whole app is one static HTML file with no build step: markup, styles and script live together in index.html. When you submit a proof, the page shrinks the photo to 1280 pixels on its longest side and sends it to Claude with the quest and a short rulebook. Claude replies with a pass or fail and a one-sentence reason, which the page shows on the quest card. The judge is told to ignore any instructions written inside the photo. One limit to know about: the judge runs on Claude's built-in features, so it works when the app is opened inside Claude. The Render version loads and saves progress, but it cannot verify photos yet. Moving the judge to a small backend that calls the Claude API is the next step. Try it Play it at touchdegrass.onrender.com and read the code on GitHub. Pull requests are welcome, especially new quests and a real backend for the photo judge. Now close the tab and go touch some grass.

VivaMate: A Local AI Interview Partner I Built for a Friend

Maltheshhh — Sun, 04 Oct 2026 05:12:39 +0000

This is a submission for the Hacktoberfest Weekend Challenge: Build for a Friend

What I Built

[Friend's name] has campus placements coming up and kept saying the same thing: "I know the answers, I just freeze when someone asks." Reading notes doesn't fix that. Practicing out loud with someone who pushes back does.

So I built VivaMate, an AI interview partner that:

Runs mock, HR, DSA and project deep-dive interviews, one question at a time
Reads [Friend's name]'s resume PDF and GitHub repos, so the questions are about their projects
Scores every answer for correctness and clarity, then tracks weak topics, streaks and a readiness score
Takes spoken answers through the mic, because freezing happens out loud
Has a 60-second Pressure Mode that simulates a real interview
Exports a PDF session report with strengths, weaknesses and revision topics

Demo

🔗 Live: https://vivamate-b7th.onrender.com

[Add a screenshot or short video of a mock interview with the readiness score visible.]

Code

maltheshhh / VivaMate

Ai based Mate for ideal viva and interviews

VivaMate

Your AI-powered companion for viva, interview, and placement preparation.

VivaMate is a local-first AI interview preparation assistant designed to help students practice technical interviews, HR rounds, DSA questions, project discussions, and resume-based interviews.

It combines conversational AI, voice input, personalized candidate context, performance scoring, readiness tracking, weak-topic detection, and downloadable reports into a single lightweight web application.

The application can run entirely locally using Ollama, making it possible to practice without sending your conversations to a cloud AI provider.

Features

AI Interview Modes

VivaMate provides multiple focused preparation modes:

Mock Interview — Practice realistic interview questions one at a time.
Explain Mode — Learn concepts you struggled with through simple explanations and examples.
HR Round — Practice behavioral and situational questions using STAR-style feedback.
DSA Round — Practice data structures and algorithms questions, including approach and complexity.
Project Deep-Dive — Discuss your GitHub projects, design decisions, trade-offs, bugs…

View on GitHub

How I Built It

The whole stack runs on open pieces:

LLM: Ollama running qwen2.5:7b, an open-weight model, with an optional OpenAI-compatible endpoint set through .env
Speech-to-text: faster-whisper, running locally
Backend: FastAPI, SQLite and streamed responses
Frontend: vanilla HTML/CSS/JS with the MediaRecorder and Speech Synthesis APIs, no framework and no CDN
Documents: pypdf to read resumes, ReportLab to write reports
Evaluation: after each answer, the model returns a structured {score, clarity, topic} JSON. That feeds SQLite and drives weak-topic detection, readiness and the /weak and /quiz commands.

Why Does Open Innovation Matter?

A mock interview is personal: your resume, your mistakes, your nervous half-answers. I didn't want a student's practice sessions sitting on a server they don't control.

Open models gave me what a closed API wouldn't:

Privacy: resume, answers and scores stay on the user's machine
Offline use: after the first model download, it works without internet
Zero running cost: no per-token bill, which matters for a student
Swappable brain: one line in .env switches the model or provider

My Agent Session

[Optional: embed or link your DevRelay session.]

Prize Categories

Best Use of Render: the live demo is hosted on Render.

My friend Arun has exams starting tomorrow. He knows the material, but reading notes doesn't prepare you for being questioned out loud, where you have to explain, defend and think under pressure.

So I built VivaMate, an AI interview and viva partner for him. It:

Asks questions one at a time and reacts to what he actually says
Reads his resume PDF and GitHub repos, so project questions are about his work
Scores every answer for correctness and clarity, then tracks weak topics, streaks and a readiness score
Takes spoken answers through the mic
Has a 60-second Pressure Mode to simulate exam-room stress
Explains concepts he gets wrong, then exports a PDF report of what to revise

devchallenge #weekendchallenge #hf26challenge