DEV Community: Priyansh Jikadra

Group-Chat-Summarizer

Priyansh Jikadra — Sun, 25 Oct 2026 16:14:22 +0000

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

What I Built

I built Group Chat Summarizer, a lazy sloth agent that reads your group chats overnight so you don't have to scroll through hundreds of messages in the morning.

You drop exported chats (WhatsApp .txt or Telegram result.json) into a folder. Every night it summarizes only the new messages in each chat into a strict two-line summary, plus a short list of action items and dates mentioned. In the morning you open one daily digest (digests/YYYY-MM-DD.md) instead of 15 noisy chats.

It's for anyone who sits in college, project, family, or society groups and misses the one important message ("class moved to 9 AM", "submit by Friday") buried under 400 memes.

Demo

🎥 Video demo: [add your link here]

Example of the morning digest in the terminal:

🦥 Group Chat Summarizer: digest for 2026-10-26

📱 CSE Batch 2027
  Most chatter was about the DSA assignment; the deadline was moved to Friday.
  Lab is shifted to Lab 3 on Monday.
  Action items: Submit DSA assignment (Fri) · Bring ID card for lab

📱 Project Team
  The team agreed to freeze features and focus on testing this week.
  Priyansh will push the final build before the demo.

Code

priyansh07-svnit / group-chat-summrizer

Hacktoberfest 2026 DEV challenge week 3 solution

group-chat-summarizer

Let a lazy sloth read your group chats overnight so you don't have to.

"Slow and steady wins the morning." 🦥

A fully local, privacy-first CLI tool that reads exported WhatsApp and Telegram group chats and produces a two-line summary of each conversation — powered by open-weight models through Ollama. No API keys, no cloud, no data leaves your machine.

Built for the DEV.to "Slow and Steady" challenge — build something with open-weight models that rewards laziness.

Demo

$ summarize run
Summarizing work_project_chat.txt (15 new messages)...
Summarizing hiking_group_chat.json (11 new messages)...

╭──────────────────────────────╮
│   Group Chat Digest          │
│       slow and steady        │
╰──────────────────────────────╯

>> Hiking Group Chat

Summary:
The group planned a Ridge Loop hike for Sunday with a 7:30 AM carpool from Main St.
Mood is enthusiastic; logistics are locked in with snacks, water, and sunscreen covered.

Action items:
- Meet 7:30 AM at Main St coffee

…

View on GitHub

How I Built It

Open-weight model: a small model (qwen2.5:3b by default) that fits comfortably on a laptop GPU with 6 GB VRAM. It's one line in config.yaml, so you can swap it for any other model.
Local inference: Ollama serves the model at localhost:11434. The app talks only to that local endpoint.
Python CLI: built with Typer and Rich. Commands: summarize run, summarize watch, summarize check, and summarize install-schedule (Windows Task Scheduler or cron).
Parsers: WhatsApp and Telegram exports are normalized into a common Message dataclass. Media placeholders and system messages are dropped.
Incremental state: a small state file remembers the last processed timestamp per chat, so each night only new messages are summarized.
Map-reduce chunking: long chats are split into chunks, each chunk is summarized, and the partial summaries are merged into the final two lines.
Editable prompt: the summary style lives in prompts/summary.txt, so you can change the tone or format without touching code.
Privacy test: a test fails if the code ever calls a non-localhost URL.
Built with Google Antigravity: I gave its agent one detailed spec and it planned, built, tested, and pushed the project to GitHub.

Why Does Open Innovation Matter?

Group chats are some of the most private text we have: family plans, friends venting, personal details. Sending them to a closed API means handing them to someone else's server. With an open-weight model, the data never leaves your laptop.

Privacy: no chat text touches a third-party server, and a test enforces it.
Works offline: after the one-time model download, the whole pipeline runs with no internet.
Zero cost: no API keys, no per-token bills, so running it every night costs nothing.
Swappable and tunable: if a model summarizes your Hinglish chats badly, change one line in the config, or fine-tune a model on your own style. A closed API wouldn't let me do either.
Editable behavior: the prompt and the code are fully mine to change.

My Agent Session

[Add your DevRelay session embed or link here]

The Lazy Payoff

Before: catching up on my groups meant scrolling through roughly [X] messages a day, about [Y] minutes.
After: I read one digest in about 30 seconds, and the nightly run needs zero effort from me.

[Replace X and Y with your real numbers from this week, or delete this section if you didn't use it.]

Prize Categories

[List the partner categories you're entering from the challenge page, or remove this section.]

Invoice Sorter

Priyansh Jikadra — Sun, 18 Oct 2026 16:45:25 +0000

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle

What I Built

Invoice Sorter is a local, privacy-first command-line tool that turns a messy folder of invoices and receipts (PDFs and images) into organized, accounting-ready data. It reads each document with OCR, uses a local open-weight LLM to pull out the vendor, amount, date, invoice number and expense category, then sorts the files into folders and exports JSON, CSV and Excel.

It's built for freelancers, solo consultants and small bookkeeping practices. They handle dozens of documents a month, and the usual options are cloud tools that upload client financial data or monthly subscriptions that eat into thin margins. Invoice Sorter runs on their own laptop, works offline, and costs nothing per document. Anything the model can't read confidently is flagged needs_review instead of being silently guessed.

Demo

Code

The full source is open under the MIT license, and the repository is embedded below. It has a src/ layout and a 49-test pytest suite.

priyansh07-svnit / invoice-sorter

project for hacktober week 2

Invoice Sorter 🧾

A local, privacy-first invoice and receipt organizer for freelancers — powered by open-source AI.

🎯 Why This Exists

Freelancers deal with dozens of invoices and receipts every month. Existing solutions either:

Send your client data to the cloud (privacy nightmare)
Cost monthly subscriptions (eats into thin margins)
Don't work offline (useless on flights or secure environments)

Invoice Sorter runs 100% locally on your laptop using Ollama and open-weight models like Llama 3.2. Your financial documents never leave your machine.

✨ Features

Feature	Description
🔒 100% Local	No cloud APIs, no data leaves your machine
🤖 Open-Source AI	Uses Ollama with Llama 3.2 (or any model you choose)
📄 Multi-Format	Processes PDFs, PNG, JPG, TIFF, BMP
🏷️ Smart Extraction	Vendor, amount, date, invoice number, category
📁 Auto-Organization	Sort by category, date, vendor, or document type
📊 Export	JSON, CSV, Excel for accounting software
💰 Free Forever	No subscriptions,

…

View on GitHub

How I Built It

Building it: I used the open-source agent harness opencode with NVIDIA's open-weight Nemotron 3 Ultra and Nemotron 3.5 Super VL models as my coding agents. Beyond writing the code, I used them to review it, find bugs and write tests.

Running it: The tool is a Python CLI built with Click and Rich. Pydantic models validate everything the LLM returns.

Text extraction: pdfplumber reads PDFs, and Tesseract OCR (via pytesseract and Pillow) reads images. Scanned PDFs without a text layer fall back to pypdfium2 rendering plus OCR.
Extraction: The text goes to a model served locally by Ollama, llama3.2:3b by default, and the model returns structured JSON. You can swap in any Ollama model with --model.
Guardrails: Model output is often messy, so parsing is lenient. It handles markdown fences, string amounts, European decimals like 1.234,56 EUR, currency symbols and null fields. Dates are validated, and anything that can't be read is flagged for review.
Output: The sorter copies files into folders by category, date, vendor or type without overwriting duplicates, and the exporter writes JSON, CSV and Excel.

Why Does Open Innovation Matter?

Invoices and receipts are some of the most sensitive documents a freelancer holds, because they contain client names, amounts and bank details. With a closed API, every document would have to be sent to a vendor's servers, and each one would cost money per call. Open-weight models make the opposite design possible. Everything runs on the user's own machine, so no data leaves it, the tool works offline, and there are no per-document costs or rate limits.

It also means users choose the model. They can run a small, fast model on a modest laptop or a larger one for better accuracy, and a model update never changes their pricing. The same openness applied to how I built the project: an open agent harness with open-weight Nemotron models meant I could build and iterate on it without being tied to a single vendor.

My Agent Session

I built this with opencode and NVIDIA's Nemotron models. [Embed your DevRelay agent_session here, or link to it.]

DEV Community: Priyansh Jikadra

Group-Chat-Summarizer

What I Built

Demo

Code

priyansh07-svnit / group-chat-summrizer

Hacktoberfest 2026 DEV challenge week 3 solution

group-chat-summarizer

Demo

How I Built It

Why Does Open Innovation Matter?

My Agent Session

The Lazy Payoff

Prize Categories

Invoice Sorter

What I Built

Demo

Code

priyansh07-svnit / invoice-sorter

project for hacktober week 2

Invoice Sorter 🧾

🎯 Why This Exists

✨ Features

How I Built It

Why Does Open Innovation Matter?

My Agent Session

My First User

Prize Categories