DEV Community: Harsh Kaushik

Ghost in the Machine: An AI That Remembers

Harsh Kaushik — Thu, 29 Oct 2026 16:33:17 +0000

An offline-first experiment in open-source AI, digital ghosts, and the stories we refuse to lose.

What I Built

What if an abandoned computer could tell stories? What if the memories we keep on our devices could live on without depending on a cloud service?

For Hacktoberfest Week 4, I’m exploring Ghost in the Machine: a local AI experience that gives forgotten digital artifacts a voice.

Inspired by Halloween and Día de los Muertos, the project explores two kinds of ghosts: the eerie personality hiding inside an old terminal and the memories of people, places, and moments that deserve to be remembered.

The goal is to create an experience that feels atmospheric and personal while keeping the underlying AI under the user's control.

Code

harsh4944 / Tic-Tac-Toe

Tic-Tac-Toe

View on GitHub

How I Built It

The project is designed around open-weight AI and local inference rather than a hosted, proprietary AI API.

The planned stack includes:

Ollama to run a compatible language model locally.
Qwen or Mistral as the conversational engine.
Python for application logic and prompt orchestration.
Streamlit for a lightweight interactive interface.

The model's personality comes from carefully designed system prompts, while the application controls how user input is processed and how responses are presented.

The architecture is intentionally modular. The model can be replaced, the prompts can be modified, and the interface can evolve without rebuilding the entire system around one vendor's API.

Why Does Open Innovation Matter?

For a project about memories and digital ghosts, ownership matters.

A closed AI service can be convenient, but it introduces dependencies: internet access, service availability, API costs, and policies controlled by someone else.

Building around open-weight models and local inference offers a different set of possibilities.

Privacy by design: Personal stories and family memories can stay on the user's machine instead of being sent to an external inference service.

Offline potential: Once the model and dependencies are installed, the core experience can work without an internet connection, assuming no online services are required by the application.

Freedom to experiment: Open model weights and configurable prompts make it possible to change the personality, compare models, and adapt the experience to different creative uses.

Less vendor lock-in: The project can evolve with its community instead of depending entirely on one proprietary endpoint.

Open innovation isn't just a way to reduce costs. It gives developers more control over how an AI behaves, where it runs, and who gets to use it.

For this project, that control is part of the point.

The Spirit of It

Halloween asks us to imagine what might be hiding in the machine. Día de los Muertos reminds us that remembering is itself an act of keeping something alive.

I wanted to explore both ideas through AI: not just a model that generates spooky text, but a tool whose architecture gives people greater control over the stories they share with it.

The most interesting part isn't whether the machine sounds haunted.

It's whether we can build something that remembers without asking us to give up ownership of what matters.

SlothDesk: A Local AI Assistant That Does the Busywork While You Rest

Harsh Kaushik — Sat, 24 Oct 2026 05:44:41 +0000

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

title: "SlothDesk: Local AI for Less Busywork"
published: false

tags: devchallenge, hf26challenge, opensource

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

What I Built

🦥 SlothDesk aims to automate repetitive tasks like summarizing notes and extracting action items, helping people save time and effort.

Demo

Coming soon.

Code

💻 [GitHub Repository — https://github.com/harsh4944/Tic-Tac-Toe]

How I Built It

The goal is to use Python and locally running open-weight AI models to summarize information and reduce manual work.

Why Does Open Innovation Matter?

Local AI can keep personal data on your device, reduce dependence on paid APIs, and let developers customize models and workflows.

The Lazy Payoff

Less repetitive work, fewer manual steps, and more time for what matters.

🦥 Slow down. Automate the busywork.

LocalLedger AI: A Private, Local-First Bookkeeper for Freelancers

Harsh Kaushik — Sat, 17 Oct 2026 14:00:29 +0000

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

What I Built

LocalLedger AI is a local-first AI assistant that helps freelancers organize invoices and receipts without sending sensitive financial documents to a third-party AI service.

Freelancers often waste time manually entering invoice numbers, dates, vendor names, tax amounts, and expense categories into spreadsheets. LocalLedger AI automates much of that repetitive work.

The workflow is simple:

Upload an invoice or receipt.
Extract relevant information using an open-weight AI model.
Review and correct the extracted fields.
Categorize expenses and export the results to CSV.

The goal is to make basic bookkeeping less tedious for independent workers who cannot justify paying for a complicated accounting platform.

Demo

[Add a link to the working demo or a short screen recording.]

Code

harsh4944 / Tic-Tac-Toe

Tic-Tac-Toe

View on GitHub

How I Built It

LocalLedger AI is designed around open-weight models and local inference rather than a proprietary AI API.

The application combines document text extraction, structured data processing, and an open-weight language model to turn unstructured invoice and receipt content into consistent records.

Potential components include:

Ollama: Run a compatible open-weight model locally.
Qwen or another suitable open-weight model: Extract and normalize invoice fields into structured JSON.
Python: Handle document processing, validation, and CSV export.
Streamlit: Provide a simple interface for uploading documents and reviewing results.
SQLite: Store extracted records locally.

The exact model and document-processing pipeline can be changed as the project evolves. Extracted financial values are validated where possible, and users can review results before exporting them.

Why Does Open Innovation Matter?

Privacy is central to this project. Invoices and receipts can contain client names, addresses, transaction details, and other sensitive information. A local-first architecture gives users control over where those documents are processed and stored.

Open-weight models also make the economics more attractive for a small side hustle. Instead of paying for a hosted inference API on every document, users can run supported models on their own hardware, subject to its capabilities and energy costs.

Open innovation provides several practical advantages:

Data ownership: Documents can remain on the user's machine.
Control over costs: Local inference avoids per-request API charges, although hardware and electricity are not free.
Model flexibility: A user can replace the underlying model as better options become available.
Customization: The extraction pipeline can be adapted to specific invoice formats, regional tax fields, and recurring business workflows.
Less vendor lock-in: The application does not have to depend on a single AI provider's pricing or API availability.

A closed API might offer stronger results on some documents or require less setup. The open approach is valuable when privacy, customization, predictable running costs, and user control matter more than outsourcing all the infrastructure.

My First User

[Add your real first-user experience here: who tried the tool, what task they completed, what feedback they gave, and whether they would pay for it.]

I want to validate whether freelancers find the time saved and privacy benefits valuable enough to pay for a polished version.

Prize Categories

[Add the partner prize categories that apply to the completed project, based on the challenge rules.]

Personalized Guidance Chatbot

Harsh Kaushik — Fri, 09 Oct 2026 06:17:37 +0000

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

What I Built

I built a Personalized Guidance Chatbot that helps users discover suitable career options based on their interests and skills. It uses AI, semantic search, and vector embeddings to provide relevant career suggestions.

In the future, I plan to extend it with outdoor learning activities and real-world skill development challenges to encourage users to move beyond the screen.

Demo

The application can be run locally using Streamlit. A live demo will be added when available.

Code

GitHub Repository: https://github.com/harsh4944/Personalized-Guidance-Chatbot

How I Built It

I used Python, Hugging Face's all-MiniLM-L6-v2 embeddings, Chroma vector database, LangChain, GPT-2, and Streamlit.

The embedding model converts user input into vectors, Chroma retrieves relevant career information, LangChain manages the retrieval process, and GPT-2 generates career suggestions. Streamlit provides the interactive chatbot interface.

Why Does Open Innovation Matter?

Open innovation allows developers to experiment with AI models, customize tools, and build useful applications without depending entirely on proprietary systems. Open-source technologies make this project easier to understand, improve, and extend with new features.

My Agent Session

Agent session link will be added if available.

Prize Categories

To be determined based on the applicable challenge partner categories.>

Study Buddy

Harsh Kaushik — Sat, 03 Oct 2026 19:25:47 +0000

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

⚡ StudyBuddy: An Aesthetic All-in-One Study Companion Built for My Best Friend

What I Built

Every exam season, I watch my friend struggle with digital clutter and burnout. Like many students, they were juggling three or four fragmented tools at once:

A separate tab for a Pomodoro timer (filled with banner ads and subscription popups),
A clunky flashcard app with paywalled decks, and
A rigid quiz tool that didn't allow quick question customization without paying for "premium tiers."

This constant context-switching caused cognitive overload and drained focus right before critical study sessions.

To solve this, I built StudyBuddy — a streamlined, distraction-free study companion designed to bring all their study workflows under one elegant, calm roof. It is built around a lightweight, zero-bloat 3-file architecture (app.py, index.html, style.css) that anyone can run locally in seconds.

🌟 Core Capabilities

⏱️ Pomodoro Focus Timer:
- Smooth circular SVG progress ring with real-time countdown.
- 3 focused modes: Focus (25m), Short Break (5m), and Long Break (15m).
- Synthesized Web Audio API bell chimes that sound without needing external audio downloads.
- Session cycle tracking (Cycle 1 of 4) and daily focus minute counters.
🗂️ Interactive 3D Flashcards:
- Realistic 3D card-flip animations on click or by pressing Space.
- Custom deck creator to organize study materials by subject.
- Spaced-repetition rating buttons (Need Practice / Mastered) to review tricky concepts.
- ⚡ One-Click Quiz Generator: Converts any flashcard deck into an interactive multiple-choice quiz instantly!
📝 Customizable Quiz System ("Where You Can Set It"):
- Lets students create their own custom quizzes tailored to their exact exam syllabus.
- Full question builder: Question prompt, 4 choices, radio selector for the correct answer, and an optional explanation note.
- Instant visual feedback (green glow on correct answer, red highlight with shake on incorrect choice) and instant rationale display.
- Results scorecard with percentage score and answer breakdown review.
🎨 Calming Aesthetic Design:
- Minimalist glassmorphic dark mode (plus toggleable crisp light mode).
- Zero external library dependencies — works offline and saves all progress directly to browser localStorage.

Demo

** Live Demo**:

harsh4944.github.io

📸 Application Preview

Pomodoro View: Circular SVG timer with glowing accent progress, customizable intervals, and session counts.
Flashcard View: 3D card scene with smooth perspective flips, keyboard shortcuts, and deck switcher.
Quiz Maker View: Clean question builder with instant feedback checks and scorecard summaries.

Code

The complete source code is open source and hosted on GitHub:

harsh4944 / studybuddy

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

⚡ StudyBuddy: An Aesthetic All-in-One Study Companion Built for My Best Friend

What I Built

Every exam season, I watch my friend struggle with digital clutter and burnout. Like many students, they were juggling three or four fragmented tools at once:

A separate tab for a Pomodoro timer (filled with banner ads and subscription popups),
A clunky flashcard app with paywalled decks, and
A rigid quiz tool that didn't allow quick question customization without paying for "premium tiers."

This constant context-switching caused cognitive overload and drained focus right before critical study sessions.

View on GitHub

👉 Direct Repository Link: https://github.com/harsh4944/studybuddy

Minimalist 3-File Architecture

studybuddy/
├── app.py          # Python Flask backend (serves app & launches browser)
├── index.html      # Responsive UI structure & self-contained client logic
├── style.css       # Complete modern CSS design system & animations
└── README.md       # Comprehensive documentation & setup instructions

How I Built It

To build StudyBuddy rapidly without sacrificing visual quality or performance, I paired with the Google Antigravity IDE autonomous coding agent:

Rapid Architecture Iteration:
- We initially explored a multi-module client-side setup and then refactored it into an ultra-clean 3-file setup (app.py, index.html, style.css) to ensure maximum simplicity, readability, and zero installation friction for my friend.
Zero-Dependency Web Audio Engineering:
- Instead of linking unreliable external .mp3 files that fail offline or hit CORS issues, we leveraged the browser's native Web Audio API to synthesize harmonic chime frequencies (C5, E5, G5, C6) and soft tactile clicks purely through code oscillators.
Pure Vanilla CSS Design System:
- Handcrafted custom CSS variables, glassmorphic backdrop blurs, 3D perspective transforms (rotateY(180deg) with transform-style: preserve-3d), and fluid CSS Grid/Flexbox layouts without heavy CSS frameworks like Tailwind or Bootstrap.
Python Flask Backend:
- Built a lightweight server in app.py that serves the application, binds to 0.0.0.0:5000, and includes an automatic browser launcher thread alongside a built-in fallback to Python’s http.server if Flask is absent.

Why Does Open Innovation Matter?

Education is the ultimate equalizer, but modern educational tools are increasingly plagued by aggressive monetization, paywalls on flashcard decks, and data-harvesting trackers.

Open innovation matters because:

Privacy & Focus: A study app should respect a student’s attention. By keeping StudyBuddy open source and offline-first, no personal data, study habits, or test scores ever leave the student’s machine.
Full Customizability: No two learners study the same way. An open codebase means my friend (or any student worldwide) can fork the repository, tweak timer durations, craft custom quiz styles, and adapt the tool to their unique learning needs.
Community Empowerment: Open source empowers developers to share educational resources freely, helping friends, classmates, and study groups grow together.

My Agent Session

This project was built with the assistance of the Google Antigravity IDE:

Pair Programming: We brainstormed the user flow, designed the Pomodoro and Quiz state machines, refined the 3D card animation physics, and verified HTTP endpoints via automated Python test harnesses.
Continuous Refactoring: When simplifying the project layout, the agent automatically removed intermediate scaffold files and unified the codebase into 3 clean, maintainable files.

Prize Categories

Build for a Friend
Most Useful Hack
Best Overall Project

Made with ❤️ for students everywhere. Happy Hacktoberfest!