DEV Community: Prabhjot Singh

FirstPR AI: Your Local AI Mentor for Your First Open-Source Pull Request 🤖

Prabhjot Singh — Tue, 20 Oct 2026 03:58:50 +0000

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

What I Built

Your first open-source pull request shouldn't feel like solving a mystery.

When I started exploring open source, I realized that contributing isn't just about writing code. Beginners must understand unfamiliar GitHub issues, navigate large codebases, figure out what to test, and prepare a pull request that follows the project's expectations. That learning curve can be intimidating.

FirstPR AI is a local AI mentor designed to make that journey easier.

Instead of spending hours trying to figure out where to begin, developers can choose a focused skill and get structured, beginner-friendly guidance.

FirstPR AI includes five AI-assisted skills:

🧩 Issue Analyzer: Breaks down an issue into its requirements, unknowns, and investigation steps.

🧭 Codebase Explorer: Helps plan how to navigate an unfamiliar repository.

🧪 Test Planner: Suggests test cases and verification steps.

🔀 PR Mentor: Guides contributors through preparing a clear, reviewable pull request.

🛡️ Contribution Safety Checker: Helps identify things to review before making changes or contributing.

The goal isn't to make developers stop learning. It's to remove the unnecessary friction that keeps beginners from getting started.

Demo

🌐 Project website: https://firstpr-ai.vercel.app/

🖥️ Run it locally: Clone the repository, install the dependencies, start Ollama, and run python app.py.

The local interface runs at http://127.0.0.1:7860.

Code

GitHub repository: https://github.com/prabhtheone/FirstPR-AI

The project is built with Python, Gradio, Ollama, and focused skill instructions. It includes automated tests and a GitHub Actions workflow for basic regression checks.

How I Built It

I built FirstPR AI around a simple idea: AI should make open source more approachable without forcing every contributor to depend on a hosted AI service.

The application uses Python and Gradio for its interface and Ollama for local model inference. Each skill has its own instructions, helping the model follow a specific workflow rather than answering every request in the same generic way.

For the Gemma 3 4B setup, the intended local model is gemma3:4b, which can be downloaded through Ollama. After the model and application dependencies are installed, inference can run on the user's own computer without requiring a cloud AI API key.

The application also includes basic error handling, input validation, and instructions that tell the model not to pretend it has inspected files, executed tests, or modified code when it hasn't.

This matters because AI-generated guidance is only a starting point. Developers must still inspect the real code, verify suggestions, run tests, and follow the repository's contribution guidelines.

Why Does Open Innovation Matter?

For a tool designed to help open-source contributors, using open technology is a natural fit.

With a local-first design, developers can experiment with AI-assisted workflows without sending issue descriptions to a third-party AI API by default. They can inspect the code, change the instructions, experiment with different open-weight models, and adapt the tool to their own workflow.

There is also a practical benefit: once the dependencies and model are installed, the core AI workflow can run locally without ongoing hosted-model API charges.

Open innovation makes the tool more than an interface around a remote service. Developers can understand how it works, improve it, and contribute back.

My Agent Session

The core workflow is available in the repository. I haven't included a separately recorded agent session in this submission.

The Lazy Payoff 🦥

The most useful kind of laziness is removing repetitive work without skipping the important parts.

FirstPR AI brings issue analysis, repository exploration planning, test planning, pull-request preparation, and contribution safety guidance into one interface.

Instead of repeatedly searching for separate tutorials and trying to connect the steps, a beginner can start with one issue and receive a structured plan.

I haven't established a measured time-saving benchmark yet, so I'm not claiming a specific number of minutes saved. The goal is to make the first step into open source less intimidating and easier to repeat.

Prize Categories

Open-Source AI Challenge — overall submission.

hf26challenge #devchallenge #opensource #ai #python #ollama #gemma #firstpr

♻️ CivicBin AI: Building an AI-Powered Waste Management Solution with Gemma 4 | From Idea to Customer Pitch

Prabhjot Singh — Tue, 13 Oct 2026 03:15:32 +0000

this is a submission for Hacktoberfest Open-Source AI Challenge — Week 2: Side Hustle

♻️ CivicBin AI: Using Gemma 4 to Tackle Real-World Waste Management

Could AI help us understand waste-disposal incidents rather than humans combing through hours of CCTV footage?

That question inspired us to create CivicBin AI — an AI-powered civic-tech project exploring how multimodal AI could help us identify potential waste-disposal issues and help us address them.

This is our GitHub repository: https://github.com/prabhtheone/CivicBin-AI

💡 The problem we were interested in solving

Waste management can be a daily occurance for communities: waste can miss bins, litter can be left on the floor and CCTV footage can take considerable time to go through before any action is taken.

We were interested in exploring whether AI could aid reviewers to understand what happened during a waste disposal event and help highlight potential incidents that may require attention.

CivicBin AI explores that area, specifically focusing on potential missed disposal — waste missing bins or being found outside of them (even though they may not be full)

🧠 Why Gemma 4?

One of the most exciting parts of this project was exploring Gemma 4 for multimodal vision-based AI

Rather than thinking of AI as a chatbot, we were interested in investigating how a vision-based model could help us interpret video footage and develop a useful application.

Gemma 4 plays a big role in our project's AI direction — our intended workflow is to provide relevant video frames, analyse the visual context and use the resulting observations to help classify a potential incident.

The important thing was connecting the model to a meaningful problem.

We also didn't want to build something just because AI is trendy — we were interested in exploring how open-model AI could contribute to a useful civic solution.

⚙️ How CivicBin AI works

Our intended workflow for CivicBin AI combines video processing with AI analysis and human review:

Video input: Start with (authorized) CCTV footage or a sample clip
Frame extraction: Extract relevant frames surrounding a possible waste event
Gemma 4 vision: Use the model to interpret the visual context
Civic Waste Investigator: Apply evidence-based rules to assess the potential incident
Structured reporting: Structure the possible incident, supporting evidence and uncertainty
Human verification: Allow a reviewer to assess the evidence before progressing

Potential outcomes include correct disposal, a missed disposal, a bin overflow or an uncertain event that requires further verification.

The aim of CivicBin AI is to assist human decision-making and shouldn't be used to directly accuse or penalise anyone.

🛠️ Building beyond the model

Working on CivicBin AI reminded us that choosing the model we were going to use was only a part of developing a useful application.

A useful AI product also includes a well-defined problem, a meaningful workflow, an approach to uncertain results and an understandable user experience.

We've also documented the project's architecture and the Civic Waste Investigator Agent Skill — which defines how a candidate incident should be reviewed, how evidence should be handled and when an event should be kept uncertain.

That way we can have a clearer direction for developing and validating the complete workflow.

🎯 Our first potential customer pitch

One of the most valuable learnings we had was considering how to pitch CivicBin AI to our first potential customer.

Building something for ourselves is one thing — selling it is another!

We had to think about it in different ways:

What problem are we solving?
Who could benefit from this solution?
How could AI-assisted incident review help their process?
Why would anyone want what we've built?

We learned that thinking as a product builder was more useful than thinking as someone that just had to complete a coding project.

We wanted to show the value of our solution and not just demonstrate AI.

🚀 From idea to shipping

We're taking CivicBin AI forward with a product mindset: building around a real-world problem, documenting our technical direction, thinking about users' needs and preparing for a solution that can be used in practice.

We want to get beyond our interesting idea and start moving towards a solution that can be evaluated in a real-world scenario.

That also means validating the model's outputs, measuring incorrect predictions, documenting limitations and ensuring human reviewers remain involved.

We believe responsible AI development also comes from more than a successful demo.

🌱 What we learned

This project has helped us explore several important areas:

Open-model AI: Understanding how Gemma 4 could be used within an application-focused workflow
Computer vision: Thinking about how video frames could support evidence collection and analysis
Agent design: Defining evidence-based review rules using the Civic Waste Investigator skill
Product thinking: Thinking about how to connect technical decisions to a real-world problem
Customer communication: Learning how to explain our project to a potential customer
Documentation: Helping other developers understand our architecture, goals and limitations
Responsible AI: Thinking about uncertainty, privacy and the need for human verification

As a first-year Computer Science student, it's taught me that development isn't just about writing code — it's about identifying problems, experimenting with technology, communicating our ideas, and improving through practical experience.

🔗 Explore CivicBin AI

GitHub Repository: https://github.com/prabhtheone/CivicBin-AI

The repository includes our project's README, architecture documentation, Gemma 4 workflow notes and Civic Waste Investigator skill.

We'd love to hear any feedback, suggestions or discussions from developers that are working with open-source AI, computer vision and civic-tech applications.

🙌 Final thoughts

CivicBin AI has been an opportunity for us to explore how Gemma 4 and open-model AI could be used for a real-world civic challenge.

From identifying the problem to thinking about our first customer pitch and the path to shipping, this has taught us to think about AI development from both technical and product perspectives.

We're excited to continue working on the project, validate the workflow and explore how AI can help us build smarter solutions for everyday challenges.

Build with purpose. Learn by doing. Ship with responsibility. ♻️🚀

Hacktoberfest #DevChallenge #HF26Challenge #Gemma4 #OpenSourceAI #ComputerVision #CivicTech #ArtificialIntelligence #BuildInPublic

🌿 Flora Scout: Offline Trail Flora & Plant Identifier

Prabhjot Singh — Mon, 05 Oct 2026 20:00:00 +0000

🌟 What I Built
Flora Scout is an offline-first mobile web application that allows hikers and outdoor enthusiasts to identify native plant species, flowers, and trees along the trail—without requiring an internet or cellular connection.

It runs an open-weight vision model locally on-device using WebGPU/Ollama, paired with an open-source agent framework (Mastra) to organize local trail findings, cross-reference native/invasive species databases locally, and log discovered flora. The app is built so that screen time is minimized to a quick 5-second snapshot, letting you get back to enjoying nature.

🍃 How It Gets People Outside ("Touch Grass")
Instead of gluing users to a screen, Flora Scout is built to enhance outdoor exploration:

Fast Trail Identification: Snap a photo on the trail, get immediate identification and brief ecology insights, and immediately put the phone away.

Zero Signal Needed: Designed specifically for dead zones in state parks and deep wilderness trails.

Gamified Local Exploration: Keeps a offline journal of indigenous species spotted during hikes, encouraging users to hit new trails and discover local ecosystems.

🔓 Why Open Innovation Matters Here
Zero Connectivity Reliance: Closed APIs (like OpenAI or Claude) fail the second you enter a forest or mountain trail with no cell signal. By running open-weight models (e.g., Gemma 2 / LLaVA) locally on-device via local inference, the app works 100% offline.

Privacy & Data Ownership: Your location data and hike photos remain strictly on your local device—never sent to cloud servers or used to train proprietary models.

Zero Operating Cost: Running open-source models directly on hardware eliminates per-token API costs completely.

🛠️ Tech Stack & Partner Tools Used
AI Core: Open-weight vision-language model (Gemma / LLaVA) for image classification and description.

Agent Framework: Mastra for agent orchestration and tool routing.

Database / Local Storage: MongoDB Atlas Edge / Local IndexedDB for caching plant taxonomy data offline.

Deployment: Render for hosting the PWA distribution server.

📸 Field Test & Demo
Took Flora Scout out to a local nature preserve with zero cellular reception. Snapped a picture of a native fern and an invasive vine on the trail—Flora Scout identified both instantly offline and flagged the invasive species with tips on safe removal!

🤖 Agent Session Record (DevRelay / Entire)
You can view the full agent session log, prompts, and debugging workflow behind building Flora Scout

FASAL AI — Intelligent Crop Classification & Quality Grading

Prabhjot Singh — Fri, 02 Oct 2026 15:50:58 +0000

FASAL AI , FASAL means crop ....
Agriculture has always been close to me because my father, friends and people around me are connected with farming. 🌾
That personal interest motivated me to try out something as simple as:

Can we make use of AI and computer vision to solve real world problems in agriculture?

And thus began my journey to create FASAL AI.

Crop AI is an open-sourced deep learning project which takes in an image of a crop and classifies it into one of five crops:

🌾 Wheat

🌾 Rice

🌽 Maize

🍌 Banana

🍈 Guava

For the Guava, it also grades the quality of the fruit into:

A → B → C → Reject

The project is made using TensorFlow, Keras and EfficientNetB0 for image classification and is ready to be expanded upon for other agricultural solutions.

📊 Here are some results:

Crop Classification: 91.76% test accuracy

Guava Quality Grading: 78.27% test accuracy

Total Dataset: 14,194 images

Crop Classes: 5

Quality Classes: A / B / C / Reject

This project is open-sourced and hosts the trained models and an inference script.

🚀 In the future, I would like to expand upon this project and do things like:

More crop classifications

Disease detection

Quality grading for more crops

Bigger and more diverse field dataset

Camera prediction in real time

Mobile applications for agriculture

For me, this isn't just another AI project. It's an attempt to use what I'm interested in (technology) to solve a problem which effects my family and community.

Hacktoberfest #OpenSource #AI #MachineLearning #ComputerVision #Agriculture #TensorFlow #Keras #EfficientNet #CropAI