DEV Community: Ahan Ghosh

🦥 SlothSort: Channeling Our Inner Sloth with 100% Offline Open-Source AI

Ahan Ghosh — Sun, 25 Oct 2026 10:12:53 +0000

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

What I Built

Sloths have mastered something we developers often forget: you don't have to do everything yourself.

While we spend hours cleaning up our Downloads folders, renaming confusing files, and listening to long voice memos, our digital clutter keeps growing.

And let's be honest, nobody wants to spend their free time organizing files.

That's why I built SlothSort — an autonomous desktop and voice memo organizer powered by open-source AI models that can run locally on your computer.

The idea is simple: let the AI handle the boring digital housekeeping while you focus on things that actually matter.

🦥 How SlothSort Works

Drop Your Files: Put your messy documents, receipts, code snippets, and voice recordings into a folder. You can also try the built-in Lazy Sandbox in the web demo.
Hibernate & Sort: Hit the Hibernate & Sort (Lazy Mode) button and let SlothSort do the work. Its AI-powered workflow categorizes files, generates cleaner filenames, analyzes documents, and transcribes audio recordings.
Wake Up to an Organized Folder: Your files are sorted into relevant categories, filenames are cleaned up using snake_case, and you receive a Morning Digest summarizing the work.
Undo Without the Stress: Made a mistake or changed your mind? Use the rollback feature to restore files to their original locations.

What Makes It Different?

Local AI: Uses open-weight models, including SmolLM2 and Whisper, for document processing and speech transcription.
Privacy First: The local desktop workflow is designed to process files without uploading their contents to a cloud AI service.
No Paid AI API Required: No recurring cloud inference charges or API keys are needed for the local AI workflow.
Voice Memo Transcription: Turn recordings into readable text files.
Morning Digest: Get a summary of what was organized, with an audio briefing option.
One-Click Rollback: Keep control over your files even when automation makes a decision you want to reverse.

The goal is to make file management something you barely have to think about.

Demo

🌐 Live Demo: https://slothsort.vercel.app/

Try the interactive showcase to explore the file organizer, Morning Digest, voice memo transcription, and rollback history.

💻 Local Desktop Mode: Run the AI workflow on your own computer, using a compatible NVIDIA GPU with CUDA acceleration or a CPU.

The live website is an interactive showcase; the local desktop application is where the offline AI processing happens.

Code

💻 GitHub Repository: https://github.com/Krishna5567/SlothSort

Explore the source code, implementation, and project architecture.

The project brings together local AI models, a Python backend, speech transcription, and a simple interface to make digital organization less tedious.

How I Built It

SlothSort combines lightweight open-weight AI models with a local-first application architecture.

The AI Stack

OpenAI Whisper: Used for automatic speech recognition, turning voice recordings into text without requiring a remote transcription API.
Hugging Face SmolLM2-360M-Instruct: A lightweight open-weight language model used for document categorization, filename suggestions, and Morning Digest generation.
pyttsx3: Supports offline text-to-speech generation for the desktop audio briefing.
Web Speech API: Provides browser-based speech synthesis in the web experience.

The Application Stack

Python: Core desktop automation and local AI workflows.
FastAPI and Uvicorn: Backend API and request handling.
PyTorch and CUDA: Support GPU-accelerated inference on compatible NVIDIA hardware.
Vercel: Hosts the interactive web showcase.

The application is designed to keep the local AI processing separate from the hosted demonstration. The website lets people explore the concept, while the desktop workflow is intended to handle their actual files locally.

Why Does Open Innovation Matter?

For a project like SlothSort, open innovation isn't just about using free software. It's about giving people more control over their data, tools, and computing resources.

🔒 1. Privacy Shouldn't Be a Premium Feature

A folder can contain bank receipts, personal notes, private recordings, and work documents. Uploading all of that to an external AI service just to organize it isn't always a good trade-off.

Running the AI locally means the desktop workflow can process sensitive files on the user's own machine instead of sending their contents to a remote AI provider.

🌐 2. Useful AI Shouldn't Depend on a Perfect Internet Connection

Cloud AI services depend on network access and external infrastructure. A local workflow can continue processing files without an internet connection once the necessary models and dependencies are installed.

That makes SlothSort useful even when the Wi-Fi isn't cooperating.

💸 3. No Per-Request AI Bill

Open-weight models let users run inference on their own hardware without paying a third-party AI provider for every request.

There are still hardware, electricity, storage, and setup costs, but the local AI workflow doesn't require a recurring cloud inference subscription.

🛠️ 4. Freedom to Experiment and Improve

Open models make it possible to inspect the available implementations, adapt prompts, experiment with alternative models, and improve the workflow without depending entirely on one closed AI provider.

That flexibility matters for developers who want to learn, contribute, and build on top of existing work.

Open innovation makes AI more accessible while giving developers and users greater control over how it works.

My Agent Session

SlothSort was developed through AI-assisted coding and an iterative build-and-test workflow.

The repository contains the project source and implementation:

Project Repository: https://github.com/Krishna5567/SlothSort
Live Showcase: https://slothsort.vercel.app/

The local workflow was tested on an NVIDIA GeForce RTX 3050 Laptop GPU with CUDA acceleration, with CPU execution also supported.

Note: I haven't included a DevRelay session recording because a shareable session link wasn't provided. Add one here if you have it.

The Lazy Payoff

The best automation is the kind that gives you your time back.

During testing, I used SlothSort on a cluttered directory containing messy filenames, documents, scripts, and voice recordings.

Instead of manually opening files, renaming them, moving them into folders, and listening through recordings, I could let the automated workflow handle those repetitive tasks.

The estimated payoff from that test was:

35–45 minutes: Estimated manual sorting and review effort avoided.
One click: Start the sorting workflow.
Readable transcripts: Get text versions of voice recordings.
One-click rollback: Restore files when needed.

These are estimates from my testing scenario, not a guarantee for every folder or computer. Actual processing time depends on the number and types of files and the available hardware.

The bigger win isn't just the time saved. It's removing one more boring task from the list of things you keep postponing.

Less digital clutter. Less repetitive work. More time to do absolutely anything else.

That's the SlothSort philosophy. 🦥

Prize Categories

Primary Challenge: Hacktoberfest Open-Source AI Challenge Week 3 — Slow and Steady.
Project Focus: Open-source AI, local inference, desktop automation, and voice memo organization.
Challenge Themes: Rewarding laziness, reducing repetitive work, protecting personal data through local processing, and making AI useful without mandatory cloud inference costs.

SlothSort is my attempt to prove that sometimes the smartest thing you can do is let your computer do the boring work while you take it slow.

Stay lazy. Stay organized. Let the sloth handle it. 🦥

FreelancerAI — The 100% Local-First AI Receipt & Expense Sorter for Solopreneurs

Ahan Ghosh — Tue, 13 Oct 2026 14:14:34 +0000

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

What I Built

FreelancerAI is a 100% local-first receipt and invoice sorter built for freelancers, solopreneurs, and small business owners.

Every quarter, millions of independent contractors face the same headache: sorting through piles of digital invoices, software receipts, cafe meeting slips, and hardware bills to prepare tax deductions. Existing cloud tools like Expensify ($5–$20/mo), Shoeboxed ($18–$36/mo), and QuickBooks Self-Employed ($15–$30/mo) charge recurring subscriptions and require uploading private financial documents to third-party cloud servers.

FreelancerAI solves this by running the entire OCR, parsing, categorization, and storage pipeline directly inside the user's browser:

Instant Local OCR: Drag and drop receipt images or invoice scans; text is extracted locally using WebAssembly.
Smart Categorization: Automatically identifies vendor name, date, subtotal, tax, and total, then maps the expense into standard tax categories (Software, Meals & Entertainment, Office Supplies, Travel, Equipment, Utilities, Professional Services, etc.).
Visual Expense Dashboard: Shows spending totals, average expense, category percentage breakdown, and monthly trends.
1-Click Tax Export: Generates clean CSV files (ready for Excel, Google Sheets, or tax preparers), structured JSON, or formatted tax summary reports.
Complete Privacy & Local Storage: No login, no accounts, and zero network calls. Data persists strictly in the browser's localStorage.

Who would pay for it?

The 60M+ freelancers worldwide who want to eliminate subscription fatigue. Instead of paying $150–$300/year for bloated accounting software, freelancers would pay a one-time fee ($19–$29 lifetime license) for an offline, private utility that just works.

Demo

Live Deployed App: https://freelancerai-six.vercel.app/
Offline Capable: Load the page once, disconnect from the internet, and test dropping a receipt—everything continues to process locally.

Code

ahanghosh77 / FreelancerAI

🧾 FreelancerAI — Local-First Invoice & Receipt Sorter

Your receipts, your device. Always.
Open-source AI-powered expense tracking that runs 100% in your browser. No servers, no uploads, no subscriptions.

🎯 What It Does

FreelancerAI is a local-first invoice and receipt sorter built for freelancers, solopreneurs, and small business owners who need to:

Scan receipts → Drop a photo, get structured data (vendor, date, amount, category)
Auto-categorize → AI classifies expenses into 10+ business categories
Track spending → Visual dashboard with category breakdown & monthly timeline
Export for taxes → Download CSV, JSON, or formatted tax summaries
Keep data private → Everything runs in your browser. Nothing is ever uploaded.

🏗️ Architecture

┌──────────────────────────────────────────────────────────┐
│                    YOUR BROWSER                          │
│                                                          │
│   ┌──────────┐    ┌─────────────┐    ┌───────────────┐  │
│   │  Image    │───▶│ Tesseract.js │───▶│ AI Rule-Based │  │
│   │  Upload   │    │ OCR Engine   │    │ Parser Engine │  │
│   └──────────┘    │ (WASM/WebGPU)│

…

View on GitHub

GitHub Repository: https://github.com/ahanghosh77/FreelancerAI
License: MIT License (permissive open-source)

How I Built It

FreelancerAI is built around open-source AI and local inference:

Open-Source AI Engine: We use Tesseract.js v5 (Apache 2.0 license), the JavaScript port of the open-source Tesseract OCR engine, running client-side inference via WebAssembly (WASM) in a Web Worker.
Local AI Parser & Categorization Engine: A rule-based natural language parser cleans noisy OCR text, extracts dates across multiple formats, detects currency symbols and amount values, isolates vendor names, and matches line items against a 300+ keyword taxonomy spanning 10 core business expense categories.
Zero-Backend Architecture: Built with semantic HTML5, Vanilla CSS3 (dark glassmorphism theme, CSS variables, micro-animations), and Vanilla JavaScript. No node server, no database, no cloud API tokens.

Why Does Open Innovation Matter?

Open-source AI made this project possible in ways closed APIs could never support:

Zero Marginal Costs ($0 API bills): Using proprietary vision models (like GPT-4o Vision or Google Cloud Vision) costs $0.02 to $0.05 per image. For a side hustle or free utility, processing hundreds of receipts per user would quickly result in unsustainable API bills. Running open-source OCR via WASM on the client's own CPU costs $0.00.
Absolute Privacy for Client Data: Freelancers routinely handle NDA-protected projects and confidential client expenses. Closed APIs require transmitting financial paperwork to cloud servers. Open innovation enables an air-gapped solution where data never leaves the device.
Offline Resilience: Open-weight and client-side models run anywhere—on a flight, in a coffee shop with spotty Wi-Fi, or during internet outages.
Sovereignty & Longevity: There are no API keys to expire, no rate limits, and no vendor price hikes or model deprecations. The code will work identically 5 years from today.

My Agent Session

This application was developed with the assistance of Google Antigravity AI through DevRelay agent workflows.

During the session, the agent:

Implemented and tuned the Tesseract.js WebAssembly OCR worker pipeline with real-time recognition progress tracking.
Crafted the regex and heuristic NLP parsing engine to handle noisy OCR character recognition (such as confusing 0 and O, varying date formats, and fragmented line items).
Designed the responsive dark glassmorphism interface and visual analytics dashboard with pure Vanilla CSS.
Executed automated browser UI and OCR verification tests with synthetic test receipts to validate the end-to-end user flow.

My First User

I shared FreelancerAI with Alex, an independent UI/UX designer and web developer who manages multiple client retainers.

What they did: Tested the app with 14 receipts from coffee shop client meetings, software tools (Figma, Adobe), and hardware purchases.
What they said: "I usually spend hours every quarter manually typing receipt numbers into a spreadsheet because QuickBooks is bloated and costs too much. FreelancerAI parsed my receipts in seconds without asking me to create an account or upload to someone else's server. The CSV export was clean and ready for my accountant."
Would they pay? Alex confirmed he would not pay for yet another monthly subscription, but said he would gladly pay a $25 to $30 one-time purchase on Gumroad to own it as an offline desktop utility.
What they asked for next: Multi-currency conversion (USD/EUR/GBP) and client-specific tags to separate billable client reimbursements.

Prize Categories

Overall Grand Prize: Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle (FreelancerAI was intentionally designed with a zero-server, 100% local client-side architecture to guarantee complete privacy and zero running costs, competing for the Overall Winner category.)

FreelancerAI — The 100% Local-First AI Receipt & Expense Sorter for Solopreneurs

Ahan Ghosh — Tue, 13 Oct 2026 14:07:06 +0000

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

What I Built

Every freelancer knows the dreaded "Tax Season Shoe-Box Panic."

Whether you are an independent software engineer, graphic designer, copywriter, or consultant, you spend 5 to 10 hours every month hunting down PDF invoices, coffee meeting slips, software subscription receipts, and hardware purchase dockets.

Existing solutions like Expensify ($5–$20/user/mo), Shoeboxed ($18–$36/mo), or QuickBooks Self-Employed ($15–$30/mo) come with two massive dealbreakers for solopreneurs:

Recurring Subscription Bloat: Paying $200–$400 every year just to digitize basic receipts eats directly into razor-thin freelance margins.
The Privacy Nightmare: Cloud OCR tools transmit your confidential client names, NDA-protected equipment purchases, credit card slips, and business travel itineraries to remote third-party servers.

Meet FreelancerAI

FreelancerAI is an open-source, local-first invoice and receipt sorter powered by client-side AI.

📸 Instant Scan & OCR: Drop any receipt image or invoice scan. Tesseract.js parses the image into machine-readable text directly inside the browser using WebAssembly.
🧠 Smart Expense Categorization: A deterministic NLP engine extracts the vendor name, transaction date, subtotal, tax, total amount, and classifies expenses into 10+ standard business tax categories (Software & Subscriptions, Office Supplies, Meals & Entertainment, Travel, Equipment, Utilities, Professional Services, etc.).
📊 Visual Financial Dashboard: Real-time analytics breaking down monthly spend, category percentage bars, average ticket size, and filterable receipt archives.
💾 100% Local Storage: Everything persists in the browser’s localStorage. No accounts, no database breaches, no tracking cookies.
📥 Zero-Friction Tax Export: Export clean CSVs (for Excel/Google Sheets), structured JSON, or print-ready Schedule C tax summary deduction reports with one click.

Who would pay for this?

The 60+ million freelancers and solopreneurs worldwide. Instead of an ongoing $20/month SaaS fee, FreelancerAI proves that open-source AI enables a sustainable "Pay once, own forever" ($19–$29 one-time license) utility or a self-hosted free community tool.

Demo

🌐 Live Web App: https://freelancerai-six.vercel.app/
⚡ Offline-Ready: Load it once, turn off your Wi-Fi, and it still scans, parses, categorizes, and exports without missing a beat.

TrailSense AI — Offline Nature Quest & Sensory Walk Companion Powered by Gemma 2

Ahan Ghosh — Tue, 06 Oct 2026 09:29:45 +0000

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

What I Built

I built TrailSense AI — an offline-first nature exploration companion designed to combat chronic screen fatigue among developers, students, and desk workers.

Instead of keeping people glued to a phone while outdoors, TrailSense AI is built around a single rule: make the screen the shortest part of the experience.

It takes just 30 seconds to choose your outdoor setting (Campus Quad, City Park, Forest Trail, or Neighborhood) and duration (15, 30, or 45 minutes). Local open-weight AI then synthesizes a personalized, tangible sensory micro-quest:

🍂 Foliage Quests: Compare leaf margin serrations between oak and maple species.
👂 Soundscape Quests: Filter out urban hum to isolate two distinct bird calls or wind rustles.
🪵 Tactile Quests: Inspect trunk bark textures (smooth birch lenticels vs deep furrowed oak).

Once generated, you pocket your phone. Gentle, non-intrusive 528 Hz audio chimes announce walk milestones and cue a 2-minute stillness mindfulness pause under a tree canopy. A built-in botanical reference identifies species with zero cell signal, and when you return, you log your findings in an offline nature journal.

Demo

🔗 Live Website: https://trailsense-ai.vercel.app/

Try it out live — select an outdoor environment, pick your duration, and generate a sensory nature quest. Features you can test:

Offline Nature Quest Dispatcher: Generates non-digital observational challenges tailored to your location.
Pocket Mode Timer: Hands-free walk tracking with gentle Web Audio chimes so your phone stays in your pocket.
Offline Field Guide: Built-in tree and bird identification reference that works without internet.
Nature Journal & Streaks: Log field notes and track your "Days Touched Grass" streak.

Code

ahanghosh77 / trailsense-ai

TrailSense AI 🌿🍂

Offline Nature Quest & Sensory Walk Companion
Built for Hacktoberfest 2026 Week 1: "Touch Grass" Challenge

🍃 Why TrailSense AI?

Modern developers and students spend 10+ hours a day staring at glowing screens. Even when we step outside, we often find ourselves mindlessly looking down at notifications, step counts, or social feeds.

TrailSense AI was built on a simple philosophy:

"Make your screen the shortest part of your day."

Spend 30 seconds on the screen to receive a sensory nature mission, pocket your phone, and let gentle audio chimes guide you through an immersive outdoor micro-quest.

✨ Features

🧭 Context-Aware Nature Quest Dispatcher: Choose your environment (Campus Quad, Urban Park, Forest Trail, Neighborhood) and duration (15, 30, or 45 minutes).
🧠 Powered by Google Gemma 2 (2B): Uses local open-weight model inference via Ollama to generate creative observational challenges (leaf identification, soundscape listening, bark texture inspection).
📴…

View on GitHub

🔗 GitHub Repository: https://github.com/ahanghosh77/trailsense-ai.git

Tech Stack: Semantic HTML5, Vanilla CSS3 (dark forest glassmorphism), Vanilla JavaScript, Web Audio API, and local Gemma 2 via Ollama with an offline heuristic botanical matrix.

How I Built It

Google Gemma 2 (2B) Prompt Architecture:
I structured strict JSON-schema system prompts for Google's lightweight open-weight gemma2:2b model running locally via Ollama. The model acts as a sensory naturalist, producing concrete, non-digital observational tasks that balance visual, acoustic, and tactile outdoor discovery.
Zero-Signal Offline Trail Architecture:
Because the best places to touch grass (deep woods, parks, trails) frequently lack cell signal, TrailSense combines local browser execution with local Ollama inference and a client-side botanical fallback matrix, ensuring 100% functionality without internet.
Screen Minimization:
Instead of prolonged screen interaction, all missions are concise and supported by gentle audio notifications.

Why Does Open Innovation Matter?

Open innovation is essential for what TrailSense AI does:

Trailside Independence: Proprietary cloud APIs fail the moment you lose cellular connection on a hiking trail. Open-weight models like Google's Gemma 2 run directly on personal devices with zero internet access.
Privacy: Location habits, daily walk routines, and personal nature reflections never leave your device.
Zero Cost: Open-source AI eliminates monthly subscriptions and API paywalls for students, hikers, and outdoor communities.

My Agent Session

This project was architected and built pair-programming with Google Antigravity, designing the local Gemma 2 prompt structure, the offline client-side botanical engine, and the Web Audio API pocket timer.

Prize Categories

Best Use of Gemma — TrailSense AI is built around Google's Gemma 2 (2B) open-weight model via Ollama, generating offline sensory nature quests and outdoor mindfulness challenges that empower developers and students to disconnect from screens and touch grass with zero subscription costs.

StudyPulse AI: An Offline-First Open-Weight Study Companion Built for my Classmate

Ahan Ghosh — Sun, 04 Oct 2026 08:13:05 +0000

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

What I Built

I built StudyPulse AI for my 3rd-semester computer science classmate and friend, Aarav.

Like many engineering students, Aarav is constantly buried under dense lecture notes in subjects like Operating Systems (semaphores, Coffman deadlock conditions, CPU scheduling), Data Structures & Algorithms (AVL tree rotations, Red-Black balance invariants), and Database Management Systems (ACID properties, BCNF normalization).

Whenever exams and midterms approach, Aarav struggles with:

Passive re-reading — highlighting slides over and over without active recall or self-testing.
Expensive API barriers — not wanting to pay $20/month for proprietary cloud AI subscriptions just to summarize homework notes.
Spotty campus internet — living in a university hostel where Wi-Fi drops constantly during study sessions.
Data privacy — not wanting course material, professor slide drafts, and private notes uploaded to corporate cloud silos.

StudyPulse AI is a lightweight, private, offline-first study companion that runs directly with local open-weight models (Google's Gemma 2, Llama 3.2, or Mistral via Ollama). It takes raw lecture notes and instantly converts them into:

🗂️ Active-Recall 3D Flip Flashcards with intuitive everyday analogies
🎯 Interactive Exam Practice Quizzes with instant feedback and grading
💡 ELI5 (Explain Like I'm 5) breakdowns
📌 High-Yield Exam Cheatsheets
📥 One-click Anki TSV and Markdown exports for long-term spaced repetition

What My Friend Thought

I shared the working build with Aarav over the weekend while we were studying for our upcoming Operating Systems evaluation:

"The fact that this runs completely offline on my laptop without requiring an internet connection or an API subscription is a lifesaver for studying in the library basement where our campus Wi-Fi always disconnects.

The killer feature for me is the everyday analogies on the back of the flashcards. I kept confusing the four Coffman deadlock conditions during revision, and the 4-way stop-sign intersection analogy made it immediately click. Also, being able to export the whole deck into Anki in one click saves me hours of formatting cards by hand."

Bonus points earned: I actually handed it over, and he's already using it for our midterms! 🎉

Demo

🔗 Live Website: https://studypulse-ai-psi.vercel.app/

Try it out — paste any lecture notes, click Generate, and flip through the 3D flashcards! Features you can test right now:

3D Flip Flashcards — click a card or press Space to flip, Arrow keys to navigate
Exam Quiz Mode — switch to quiz mode for interactive MCQs with instant scoring
ELI5 Analogies — every answer comes with a real-world analogy that makes concepts stick
Quick Samples — click "OS: Deadlocks", "Data Structures: Trees", or "DBMS: ACID & Normalization" to see it in action instantly
Export — download your deck as Anki-ready TSV or clean Markdown notes

Code

ahanghosh77 / studypulse-ai

⚡ StudyPulse AI

Private Open-Weight Note-to-Flashcard & Exam Prep Assistant
Built for a classmate for the Hacktoberfest 2026 Weekend Challenge: "Build for a Friend"

🌟 Overview

StudyPulse AI is an offline-first, privacy-respecting study companion that transforms dense semester lecture notes, textbook passages, and slides into active-recall flashcards, practice quizzes, intuitive everyday analogies, and Anki-ready decks using local open-weight AI models.

Built specifically for engineering and CS students tackling tough 3rd-semester subjects like Operating Systems, Data Structures & Algorithms, and Database Management Systems.

✨ Features

🗂️ 3D Active-Recall Flashcards: Spatial card flips with keyboard shortcuts (Space to flip, Arrow keys to navigate) and mastery tracking.
🎯 Exam Practice Quiz: Auto-generated multiple-choice questions with instant rationale and scoring.
💡 ELI5 & Intuitive Analogies: Translates complex system concepts into relatable mental models.
🔒 100% Offline & Private: Connects locally to Ollama (gemma2:2b, llama3.2:3b, mistral) or any local…

View on GitHub

Tech Stack:

Frontend: Semantic HTML5, Vanilla CSS3 (glassmorphism, 3D perspective card flip animations, responsive grid), Vanilla JavaScript
AI Core: Direct integration with Ollama (gemma2:2b, llama3.2:3b) and any OpenAI-compatible local inference server, backed by an offline heuristic parsing engine
Exports: Anki (.tsv) spaced repetition exporter and Markdown generator
Deployment: Vercel (static)

How I Built It

Prompt Engineering for Open-Weight Models:
Open-weight models like gemma2:2b or llama3.2:3b are efficient and run smoothly on modest student laptops. I structured strict JSON-schema system prompts requiring the model to extract:
- A core question testing causal understanding (not simple rote memorization)
- A concise, high-yield exam answer
- A relatable, real-world everyday analogy (e.g., explaining Coffman deadlock conditions like a 4-way stop sign intersection, or ACID atomicity like an ATM cash withdrawal)
Resilient Local Architecture:
The app communicates with http://localhost:11434/api/generate via browser fetch with configurable model and endpoint options. If the student hasn't started their local Ollama daemon, StudyPulse seamlessly falls back to its built-in heuristic semantic parser — guaranteeing 100% uptime whether online or offline.
Active Recall & Spaced Repetition Design:
Instead of overwhelming students with walls of text, StudyPulse focuses on atomic cards, self-evaluations ("Mastered" vs "Needs Review"), mastery progress tracking, and Anki export compatibility for long-term retention.

Why Does Open Innovation Matter?

Open innovation is what made StudyPulse possible:

Zero Cost for Students: College students shouldn't need monthly subscriptions to generate flashcards from their own notes. Open-weight models like Gemma and Llama bring state-of-the-art reasoning to personal computers for free.
Complete Privacy: Course notes, university research, and student drafts never leave the device. No telemetry, no cloud training on user data, zero data leakage.
True Offline Independence: Open-source AI doesn't depend on cloud server status pages, rate limits, or campus Wi-Fi drops. It runs reliably on an airplane, in a library basement, or in a dorm room with no internet.
Customizability: Students can fine-tune small open-weight models on their own department's curriculum or swap models freely — no vendor lock-in.

A closed API couldn't do any of this: it would cost money every month, require constant internet, upload private academic notes to remote servers, and stop working the moment the provider's rate limit kicks in or the Wi-Fi drops.

My Agent Session

This project was built pair-programming with Google Antigravity and saved via DevRelay. The session captures how we planned the open-weight offline architecture, structured the 3D CSS flip animations, and designed the active-recall study heuristics.

Session 2026-10-04 07:45

You

I want to participate in the Hacktoberfest Weekend Challenge: Build for a Friend. Let's build a Classmate Semester Study Assistant / Note-to-Flashcard generator.

Agent

Designed and built StudyPulse AI, an offline-first study companion tailored for a 3rd-semester engineering classmate tackling Operating Systems, Data Structures, and DBMS. Implemented 3D active-recall flashcards, practice exam quizzes with instant scoring, mental model analogies, Anki TSV/Markdown exports, and direct integration with local open-weight models via Ollama with zero cloud dependencies.

▸ write_to_file index.html, style.css, app.js

Input

index.html, style.css, app.js

Output

Successfully generated full modern web application with glassmorphism and 3D flip card animations.

🔗 Session Transcript: https://dev.to/agent_sessions/session-2026-10-04-0745-4fphv3

Prize Categories

Best Use of Gemma — StudyPulse AI is built to run with Google's Gemma 2 open-weight model (gemma2:2b) locally via Ollama, generating active-recall flashcards, practice quizzes, and intuitive analogies 100% offline with zero subscription cost.

StudyPulse AI: An Offline-First Open-Weight Study Companion Built for my Classmate

Ahan Ghosh — Sun, 04 Oct 2026 07:40:13 +0000

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

What I Built

I built StudyPulse AI for my 3rd-semester computer science classmate and friend, Aarav.
Like many engineering students, Aarav is constantly buried under dense lecture notes in subjects like Operating Systems (semaphores, Coffman deadlock conditions, CPU scheduling), Data Structures & Algorithms (AVL tree rotations, Red-Black balance invariants), and Database Management Systems (ACID properties, BCNF normalization).
Whenever exams and midterms approach, Aarav struggles with:

Passive re-reading: Highlighting slides over and over without active recall or self-testing.
Expensive API barriers: Not wanting to pay $20/month for proprietary cloud AI subscriptions just to summarize homework notes.
Spotty campus internet: Living in a university dorm where Wi-Fi drops constantly during study sessions.
Data privacy: Not wanting course material, professor slide drafts, and private notes uploaded to corporate cloud silos. StudyPulse AI is a lightweight, private, offline-first study companion designed to run directly with local open-weight models (such as Google’s Gemma 2, Llama 3.2, or Mistral via Ollama/LocalAI). It takes raw lecture notes or textbook snippets and instantly converts them into:
🗂️ Active-Recall 3D Flip Flashcards with intuitive everyday analogies
🎯 Interactive Exam Practice Quizzes with instant feedback and grading
💡 ELI5 (Explain Like I'm 5) breakdowns
📥 One-click Anki TSV and Markdown exports for long-term spaced repetition ### What My Friend Thought (Real User Feedback) I shared the working build with Aarav over the weekend while revising Operating Systems. Here is his feedback: > "The fact that this runs completely offline on my laptop without requiring an internet connection or an API subscription is a lifesaver for studying in the library basement where our campus Wi-Fi always disconnects. > > The killer feature for me is the everyday analogies on the back of the flashcards. I kept confusing the four Coffman deadlock conditions during revision, and the 4-way stop-sign intersection analogy made it immediately click. Also, being able to export the whole deck into Anki in one click saves me hours."

Demo

studypulse-ai-psi.vercel.app

Code

ahanghosh77 / studypulse-ai

⚡ StudyPulse AI

Private Open-Weight Note-to-Flashcard & Exam Prep Assistant
Built for a classmate for the Hacktoberfest 2026 Weekend Challenge: "Build for a Friend"

🌟 Overview

Built specifically for engineering and CS students tackling tough 3rd-semester subjects like Operating Systems, Data Structures & Algorithms, and Database Management Systems.

✨ Features

🗂️ 3D Active-Recall Flashcards: Spatial card flips with keyboard shortcuts (Space to flip, Arrow keys to navigate) and mastery tracking.
🎯 Exam Practice Quiz: Auto-generated multiple-choice questions with instant rationale and scoring.
💡 ELI5 & Intuitive Analogies: Translates complex system concepts into relatable mental models.
🔒 100% Offline & Private: Connects locally to Ollama (gemma2:2b, llama3.2:3b, mistral) or any local…

View on GitHub

How I Built It

Prompt Engineering for Open-Weight Models:

Open-weight models like gemma2:2b or llama3.2:3b are efficient and run smoothly on modest student laptops. I structured strict JSON-schema system prompts requiring the model to extract:
- A core question testing causal understanding (not simple rote memorization)
- A concise, high-yield exam answer
- A relatable, real-world everyday analogy (e.g., explaining Coffman deadlock conditions like a 4-way stop sign intersection, or ACID atomicity like an ATM cash withdrawal).
Resilient Local Architecture:

The application communicates with http://localhost:11434/api/generate via browser fetch with configurable model and endpoint options. If a student is on the go and hasn't started their local Ollama daemon, StudyPulse seamlessly falls back to its built-in heuristic semantic parser, guaranteeing 100% uptime.
Active Recall & Spaced Repetition Design:

Instead of overwhelming the student with giant walls of text, StudyPulse focuses on atomic cards, self-evaluations ("Mastered" vs "Needs Review"), and Anki export compatibility.

Why Does Open Innovation Matter?

Zero Cost for Students: College students shouldn't have to budget monthly subscriptions just to generate study flashcards from their own notes. Open-weight models like Gemma and Llama bring state-of-the-art reasoning to everyday personal computers for free.
Complete Privacy: Course notes, university research, and student drafts never leave the student's device. No telemetry, no cloud training on user notes, and zero data leakage.
True Offline Independence: Open-source AI isn't dependent on cloud server status pages, rate limits, or campus Wi-Fi connection drops. It runs reliably on an airplane, in a library basement, or in a dorm room.
Customizability: Because the models are open, students can fine-tune small open-weight models on their own department's curriculum or switch models freely without vendor lock-in.

My Agent Session

This project was developed with the assistance of Google Antigravity and configured with DevRelay for Hacktoberfest 2026. The agent workflow helped scaffold the semantic extraction prompts, structure the 3D CSS flip animations, and curate high-yield 3rd-semester computer science concepts.

Prize Categories

Best Use of Gemma: StudyPulse AI is built to run directly with Google's Gemma 2 open-weight model (gemma2:2b) locally on modest student laptops via Ollama/local inference, generating active-recall flashcards, practice quizzes, and intuitive analogies 100% offline with zero subscription cost.

New to Community

Ahan Ghosh — Sun, 04 Oct 2026 04:55:00 +0000

Hellooo, everyone@!!!