DEV Community: Rafia Minhaj

EcoPath AI: Offline Open-Weight AI Assistant for Outdoor Hikers & Nature Trails

Rafia Minhaj — Tue, 06 Oct 2026 10:07:18 +0000

🌿 The "Touch Grass" Challenge & The Offline Problem

When you go outside to hike, explore nature trails, or garden, cellular network signal drops to zero. Closed AI APIs (like OpenAI or Claude) completely break down on the trail without active internet connection.

To truly "Touch Grass" and explore the outdoors safely, hikers need an AI assistant that runs 100% locally using open-weight models without any internet connection.

⚡ Introducing EcoPath AI

EcoPath AI is an open-source, offline-first outdoor assistant powered by open-weight AI models. It helps outdoor enthusiasts identify plant species, plan safe hiking routes, and receive voice-guided nature trail alerts hands-free.

🌟 Key Features:

Zero-Signal Offline Inference: Runs open-weight models (Gemma 2B / Llama 3) locally via Ollama. Works deep in forests with zero internet!
Hands-Free Audio Trail Guidance: Integrates ElevenLabs audio synthesis to narrate trail checkpoints and plant descriptions without forcing you to look at a phone screen.
Local Route & Flora RAG: Uses Backboard RAG memory for storing local flora data and offline map embeddings.
Cloud Sync via Render: When back online, trail logs and route updates automatically sync to the cloud hosted on Render.

💡 Why Open Innovation Matters Here

Open innovation is the ONLY reason EcoPath AI exists:

Privacy & Data Sovereignty: Your real-time GPS coordinates and nature logs stay 100% on your device, off private corporate servers.
Zero Cost & No Internet Dependency: Closed APIs cost money per call and require 4G/5G signal. Open-weight models cost ₹0 to run offline on a laptop or edge device during trail walks.

🛠️ Tech Stack & Partner Technologies

Open AI Core: Open-weight LLMs (Gemma / Llama 3) via Ollama
Audio Engine: ElevenLabs Voice API (Hands-Free narration)
Agent Memory: Backboard RAG Engine
Cloud Infrastructure: Render (Backend API Runtime)

📸 Try It Out & Open Source Code

GitHub Repository: https://github.com/RafiaMinhaj
Author: Rafia Minhaj (B.Tech CSE '27 @ CIT Ranchi | GSSoC '26 Global Rank #29 | GfG Campus Lead)

Submitted to the Hacktoberfest 2026 Week 1 DEV Challenge ("Touch Grass" Theme)! #devchallenge #hf26challenge

TechPulse AI — Autonomous Engineering Interview Errata & Verification Agent

Rafia Minhaj — Wed, 23 Sep 2026 15:52:41 +0000

This is a submission for the Sanity Challenge, Path One: Ship an Agent That Queries Real Content

What I Built

I built TechPulse AI, an autonomous engineering interview verification agent designed to resolve contradictions between company coding specifications, rulebooks, and edge-case errata for top technical interview drives (such as Booking Holdings, Infosys SP, IBM, and AWS).

TechPulse AI queries structured JSON documents stored in a Sanity Content Lake via the Sanity Context MCP Endpoint, evaluating claims side-by-side with source provenance and returning authoritative, source-verified answers.

Demo

Live Web Application: https://archscale-lead-engine.vercel.app
Sanity Context Dataset: https://github.com/rafiaminhaj/techpulse-sanity-agent

Code

GitHub Repository: https://github.com/rafiaminhaj/techpulse-sanity-agent

The project architecture consists of:

Sanity Content Lake Schema: interviewQuestion, companySpec, errataDoc.
FastAPI Microservice: Async agent backend connecting to Sanity Context MCP.
React/Next.js Dashboard: Interactive query interface displaying source claims side-by-side.

How I Used Sanity

I modeled technical interview knowledge into structured JSON schemas inside Sanity:

companySpec: Tracks time limits, memory limits, and platform constraints.
errataDoc: Stores verified edge-case clarifications and rule contradictions.

Key Sanity Context Tools Used:

Sanity Context MCP Endpoint: Exposed structured datasets through MCP to our Python AI agent.
GROQ Queries: Retrieved precise schema relationships linking errata back to authoritative company specs.
Knowledge Base Indexing: Distilled technical docs into indexed, queryable chunks where every claim retains source provenance.

When a query is asked, TechPulse AI fetches both the rulebook spec and the errata doc simultaneously, surfacing conflicting constraints side-by-side with direct links to the underlying Sanity documents.

Sanity Project Details

Sanity Project ID: techpulse_ai_2026
Dataset: production

Agent Session

Agent Session Transcript: Embedded via Gemini CLI and FastAPI MCP client.

Author

Rafia Minhaj

B.Tech CSE '27 | GSSoC '26 Global Rank #29 | Microsoft Certified AI Agent Engineer

GitHub: https://github.com/rafiaminhaj

PawsitiveAI 🐾 - AI Dog Breed Classifier & Voice Companion

Rafia Minhaj — Sat, 15 Aug 2026 18:26:10 +0000

Introducing PawsitiveAI 🐾

For the DEV Weekend Challenge (Dog Days Edition), I built PawsitiveAI—a sleek, dark-themed, glassmorphic web application that serves as an AI-powered dog breed classifier and interactive voice companion.

It helps dog owners, prospective adopters, and dog lovers identify breeds from photos, discover traits, and listen to a custom voice-generated care guide!

🔗 Project Links

GitHub Repository: github.com/Rafiaminhaj/pawsitive-ai
API Technologies Used: Google Gemini API & ElevenLabs API

🚀 How It Works

Upload & Analyze: Drag and drop or browse any dog image.
Google AI Reasoning: The backend FastAPI server sends the image to Google's Gemini 1.5 Flash model, which classifies the breed and generates structured metadata (traits, care tips, fun facts, exercise requirements, and health indices).
ElevenLabs Voice Synthesis: Clicking the "Listen to Care Guide" button triggers the ElevenLabs API to speak the personalized care guide in a clean, natural voice.
Smart Fallback: If API keys are absent, the application gracefully activates Simulation Mode (mapping popular breeds) and leverages the browser's native Web Speech API for offline speech generation, ensuring a seamless, zero-cost UX for everyone.

🛠️ The Tech Stack

Backend Framework: FastAPI / Uvicorn (Python)
Frontend Design: Glassmorphic HTML5, CSS3 Variables, ES6 JavaScript (No heavy frameworks for maximum load speed).
AI Processing: Deprecating traditional vision APIs for Google AI Generative Models (google-generativeai).
Audio Synthesis: ElevenLabs TTS REST API + Browser SpeechSynthesis.

📸 Interface Preview

Here is how the PawsitiveAI dashboard looks after classifying a Golden Retriever:

Accurate Match Confidence Indicators
Dynamic Telemetry Dial Meters (Exercise & Health Indices)
Interactive Play/Pause Voice Controllers

(Check out the repository for screenshots and codebase structure!)

💡 What I Learned

During this weekend build, I learned how to handle unstructured image inputs to return strict JSON structures from LLMs (Gemini) and parse them safely on the client side. I also enjoyed designing a premium glassmorphic UI using pure CSS variables and radial gradient light sources.

Thank you to the DEV Community and the sponsors (Google AI & ElevenLabs) for organizing this challenge! 🐾

The Jharkhand Kitchen — A Heritage Culinary Landing Page

Rafia Minhaj — Sun, 09 Aug 2026 19:18:10 +0000

What I Built

I built The Jharkhand Kitchen, a premium, highly interactive frontend landing page celebrating the rich culinary heritage of Jharkhand, India. The page highlights two legendary regional comfort foods: Dhuska (a crispy, golden deep-fried rice & lentil savory bread) and Litti Chokha (rustic charcoal-baked whole wheat balls stuffed with spiced sattu, glazed with ghee).

The highlight of the project is a detailed concentric bronze plate containing the dishes, crafted entirely using pure CSS Art (gradients, box-shadows, and keyframe animations for steam).

Link to Source Code

👉 GitHub Repository: Rafiaminhaj/jharkhand-comfort-food

Key Features & Tech Stack

Pure CSS Art: No external images or SVGs were used to render the bronze plate, Dhuskas, Litti, clay bowls, or steam. Everything is styled using layered gradients, border-radius manipulation, and CSS animations.
3D Parallax Tilt Effect: Mouse movement over the food platter and cards calculates the cursor offset in JavaScript to trigger smooth 3D tilting.
The Culinary Duel: An interactive voting system that lets users vote for Team Dhuska or Team Litti Chokha, complete with LocalStorage persistence and coordinate-based custom particle confetti explosions.
Web Accessibility: Focus rings, skip-to-content links, semantic HTML, and an ARIA live region announcer for screen readers to hear vote updates.

Journey & What I Learned

Through this challenge, I pushed my CSS styling skills to the absolute limit. Designing the organic texture of fried Dhuskas and the smoky, charcoal-burned details on the Litti using only CSS gradients and box-shadows was a huge learning experience. Implementing coordinate-based particle math in vanilla JavaScript to make confetti explode from the exact button click location was highly rewarding!