DEV Community: Sanjana jha

Local Expense Sorter: an offline categorizer for freelancers and small shops

Sanjana jha — Sun, 18 Oct 2026 17:19:38 +0000

devchallenge #hf26challenge

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

What I Built

Local Expense Sorter takes a bank statement CSV and categorizes each transaction (Food & Drinks, Supplies, Rent, Client Income, and so on) using a small open-weight model running on the user's own laptop. It then shows monthly spend charts and exports an Excel workbook or CSV for an accountant.

It's for freelancers and small shop owners who spend hours sorting bank and UPI statements by hand. Most tools that do this upload financial data to the cloud. This one doesn't. When a user corrects a category, that correction is saved and used as an example for future imports.

Demo

No hosted link. Here's the dashboard after importing a sample statement:

To try it, follow the setup steps in the README and upload the sample CSV included there.

Code

https://github.com/sanjana-jha-001/expense-sorter

How I Built It

Model: Qwen2.5 3B (open weights), run locally through Ollama
Storage: SQLite, a local file. Duplicate transactions are detected with a hash, so re-importing a statement is safe.
Interface: Streamlit and pandas, bound to 127.0.0.1 so only the same machine can reach it

Categorization runs in order of precedence. A saved correction for a merchant is used first, with no model call. Repeat merchants in the same batch reuse the earlier result. Only new merchants go to the model, and the 20 most recent corrections are included in the prompt as examples.

Why Does Open Innovation Matter?

Privacy: Bank statements show client names and family spending. With a local model, nothing leaves the laptop. The README shows how to run the app with --network none to verify this.
Cost: Per-transaction API fees would eat the margin on a small paid tool. Local inference costs only electricity.
Control: I can swap models, edit the prompt, or change categories without asking a vendor.

The trade-off is accuracy. A 3B model is less accurate than a large cloud model on messy bank descriptions, so the review screen is part of the design, and every correction improves later imports.

TouchGrass AI** 🌿 — an AI-powered outdoor activity recommendation system

Sanjana jha — Thu, 08 Oct 2026 16:15:25 +0000

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 1: Touch Grass

What I Built

I built TouchGrass AI 🌿 — an AI-powered outdoor activity recommendation system designed to do something unusual: help people spend less time with technology.

Instead of recommending another app, video, or online activity, TouchGrass AI gives you a small real-world mission to complete outside.

You tell it:

🧠 Your current mood
🎯 Your goal
⏱️ How much time you have
🌿 Your preferred activity
🌳 Your environment
👥 Whether you're alone or with someone

It then recommends a personalized outdoor activity that fits your situation.

For example:

Light Hunt

Find an interesting patch of sunlight and notice three different shadows.

Or:

No-Route Walk

Take a short walk without following your usual route and discover three things you normally overlook.

The idea is simple:

Use AI for a few seconds → Get a mission → Put your phone down → Go outside.

TouchGrass AI is designed for anyone who feels like they spend too much time in front of a screen and wants a simple reason to step outside.

Demo

🌿 Try TouchGrass AI:

https://touchgrass-mmewgdojhn6dqggh52qkfs.streamlit.app/

The goal of the experience is intentionally short. Once you receive your mission, you're encouraged to stop looking at the screen and actually do it.

Code

💻 GitHub Repository:

https://github.com/sanjana-jha-001/touchgrass

The project is open source and built with Python and Streamlit.

How I Built It

TouchGrass AI is built with Python + Streamlit and combines a built-in recommendation system with optional local AI inference.

The project can use Ollama with open-weight models to generate personalized outdoor missions.

The recommendation prompt gives the model the user's:

Mood
Goal
Available time
Preferred activity
Environment
Social situation

The model is specifically instructed to recommend activities that get the user away from the screen, rather than suggesting more digital content.

I also included a built-in recommendation system, so the application can still work without a running AI model.

One important design decision was making AI an enabler rather than the destination.

The successful outcome isn't another long AI conversation.

The successful outcome is:

The user closes the app and goes outside. 🌱

Why Does Open Innovation Matter?

Open innovation is especially important for this project because I wanted the AI component to be transparent, accessible, and replaceable.

Using open-weight models and local inference through Ollama means the project doesn't have to depend entirely on a proprietary cloud AI API.

That makes it possible to:

Run AI locally
Experiment with different open-weight models
Customize the recommendation prompt
Build without requiring a paid API key
Keep the AI component flexible and replaceable
Learn how the AI component actually works

For a project whose goal is to reduce dependence on screens and online services, using local AI felt especially appropriate.

My Agent Session

I used AI-assisted development to help design, implement, debug, and improve TouchGrass AI.

The project itself is focused on using AI responsibly: instead of optimizing for more engagement, it uses AI to encourage less screen time and more real-world activity.

Prize Categories

Open-Source AI / AI-powered application
Open Innovation

Final Thought

Most technology is designed around:

"How can we keep the user engaged?"

I wanted to experiment with the opposite question:

"How can AI help the user leave?"

That's TouchGrass AI. 🌿

Use AI for a few seconds.

Get your mission.

Touch grass.

🌿 I Built TouchGrass AI — An AI That Wants You to Use Your Phone Less

Sanjana jha — Thu, 08 Oct 2026 16:11:29 +0000

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 1: Touch Grass

What I Built

I built TouchGrass AI 🌿 — an AI-powered outdoor activity recommendation system designed to do something unusual: help people spend less time with technology.

Instead of recommending another app, video, or online activity, TouchGrass AI gives you a small real-world mission to complete outside.

You tell it:

🧠 Your current mood
🎯 Your goal
⏱️ How much time you have
🌿 Your preferred activity
🌳 Your environment
👥 Whether you're alone or with someone

It then recommends a personalized outdoor activity that fits your situation.

For example:

Light Hunt

Find an interesting patch of sunlight and notice three different shadows.

Or:

No-Route Walk

Take a short walk without following your usual route and discover three things you normally overlook.

The idea is simple:

Use AI for a few seconds → Get a mission → Put your phone down → Go outside.

TouchGrass AI is designed for anyone who feels like they spend too much time in front of a screen and wants a simple reason to step outside.

Demo

🌿 Try TouchGrass AI:

https://touchgrass-mmewgdojhn6dqggh52qkfs.streamlit.app/

The goal of the experience is intentionally short. Once you receive your mission, you're encouraged to stop looking at the screen and actually do it.

Code

💻 GitHub Repository:

https://github.com/sanjana-jha-001/touchgrass

The project is open source and built with Python and Streamlit.

How I Built It

TouchGrass AI is built with Python + Streamlit and combines a built-in recommendation system with optional local AI inference.

The project can use Ollama with open-weight models to generate personalized outdoor missions.

The recommendation prompt gives the model the user's:

Mood
Goal
Available time
Preferred activity
Environment
Social situation

The model is specifically instructed to recommend activities that get the user away from the screen, rather than suggesting more digital content.

I also included a built-in recommendation system, so the application can still work without a running AI model.

One important design decision was making AI an enabler rather than the destination.

The successful outcome isn't another long AI conversation.

The successful outcome is:

The user closes the app and goes outside. 🌱

Why Does Open Innovation Matter?

Open innovation is especially important for this project because I wanted the AI component to be transparent, accessible, and replaceable.

Using open-weight models and local inference through Ollama means the project doesn't have to depend entirely on a proprietary cloud AI API.

That makes it possible to:

Run AI locally
Experiment with different open-weight models
Customize the recommendation prompt
Build without requiring a paid API key
Keep the AI component flexible and replaceable
Learn how the AI component actually works

For a project whose goal is to reduce dependence on screens and online services, using local AI felt especially appropriate.

My Agent Session

I used AI-assisted development to help design, implement, debug, and improve TouchGrass AI.

The project itself is focused on using AI responsibly: instead of optimizing for more engagement, it uses AI to encourage less screen time and more real-world activity.

Prize Categories

Open-Source AI / AI-powered application
Open Innovation

Final Thought

Most technology is designed around:

"How can we keep the user engaged?"

I wanted to experiment with the opposite question:

"How can AI help the user leave?"

That's TouchGrass AI. 🌿

Use AI for a few seconds.

Get your mission.

Touch grass.

College Copilot: an AI that tells my friend what matters today

Sanjana jha — Sun, 04 Oct 2026 12:29:53 +0000

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

What I Built

Ask any college student where their important updates live and the honest answer is: everywhere. WhatsApp groups, notice PDFs, screenshots of notice boards, assignment deadlines, internship forwards.

I built College Copilot for my friend Keerti, a B.Tech student, so she doesn't have to dig through all of that. It reads her college chaos and answers one question: "What matters for ME right now, and what should I do?"

Instead of just summarising notices, it matches them with her subjects, attendance and interests and tells her what to do:

⚠️ DBMS attendance is 72%, 75% is required. Attend the next 3 classes.

What it does:

Reads everything: WhatsApp chat exports, notice PDFs, screenshots and pasted text, including Hinglish like "kal submit karna hai"
Filters for her: ignores notices meant for other branches and ranks the rest by urgency
Attendance planner: shows how many classes she must attend in each subject, with a bunk simulator
Deadlines, exams and internships in separate tabs, with apply links
Ask anything, in Hinglish ("is hafte kya submit karna hai?")
Calendar export to Google Calendar, and a ready-to-send Hinglish morning message

Demo

Live app: https://college-copilot-jdgv5fsevwfc4jbcgzmyrz.streamlit.app/

The live app is hosted and sends text to a cloud model, so I only use made-up sample chats and notices on it. Please don't enter real private data there.

Code

sanjana-jha-001 / college-copilot

🎓 College Copilot - private, offline AI for one student

Reads scattered college info (WhatsApp exports, notice PDFs, screenshots), matches it to the student's subjects/attendance/interests and answers: "What matters for ME right now and what should I do?"

Run

Install Ollama, then: ollama pull qwen2.5:7b (optional vision: ollama pull qwen2.5vl:7b)
pip install -r requirements.txt
streamlit run app.py
Upload sample_chat.txt from the sidebar to test, then your friend's real chat.

Features

WhatsApp chat parser (Android + iOS formats), Hinglish-aware extraction ("kal", "aaj")
Notice PDFs + screenshots (local vision model)
Priority ranking by urgency x your subjects x your interests; other-branch notices are dropped
Attendance planner: classes you must attend, safe bunks left, bunk simulator (pure code, no LLM math)
Internship tab with links, Ask-anything chat (Hinglish), ready-to-send morning WhatsApp message
Calendar (.ics) export, profile saved locally

Why open source

Private chats and attendance never leave the laptop, works with…

View on GitHub

How I Built It

Open-weight model: the hosted demo runs an open-weight model (default: OpenAI's gpt-oss-120b) served through Groq. The app fetches the models Groq currently offers and lets me pick one from a dropdown.
Local option: the same LLM layer in core.py can talk to a local Ollama model instead, so the app isn't tied to one provider.
Streamlit for the interface, Python for WhatsApp parsing, PDF reading and ranking.
The AI only reads, code does the maths. The LLM extracts deadlines and actions from messy text. The attendance calculation, the "classes you must attend" count and the priority ranking are plain code, so the numbers are never hallucinated.
Hinglish dates: the prompt resolves words like "kal" and "aaj" relative to the date of each message, so deadlines get real dates.

Why Does Open Innovation Matter?

This app deals with some of the most personal data a student has: private group chats and attendance records.

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

No lock-in. While building this, Groq retired a model I had hard-coded and my app broke. Because the model is just a setting, I moved to another open model with one dropdown instead of rewriting anything.
Where the data goes is a choice. The same open model can run on a hosted service for a public demo, or on the student's own laptop through Ollama when privacy matters.
No dependency on one vendor's pricing or availability.