Walk Tutor: a voice DSA coach that runs offline on my laptop

Marturu Ramyasri — Sun, 11 Oct 2026 13:18:32 +0000

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

What I Built

Walk Tutor is a voice coach for practicing data structures and algorithms (DSA) without staring at a screen. It reads you a problem, you explain your approach out loud, and it answers with a hint or a follow-up question. You can say "next problem", "repeat" or "quit", so there is no keyboard at all once it starts.

I'm a third-year student, and most of my practice is sitting at a laptop. I built this so I could practice while walking outside instead.

Code

ramya2006sri / walk-tutor

Walk Tutor

A voice DSA coach you use on a walk. It runs fully offline on a laptop.

How it works

Speech to text: faster-whisper
Coach: Gemma 3 1B through Ollama
Speech out: pyttsx3
The code checks your answer for the key idea; the model only writes the feedback.

Setup

Install Ollama and run: ollama pull gemma3:1b
pip install -r requirements.txt
python walk_tutor.py

Voice commands: "next problem", "repeat", "quit".

View on GitHub

How I Built It

Everything runs locally on my Windows laptop:

Gemma 3 1B through Ollama writes the coach's replies.
faster-whisper (open-source Whisper) turns my speech into text.
pyttsx3 speaks the answers using the voices built into Windows.
A small Python loop detects when I start and stop talking, so it needs no button presses.

The lesson that shaped the design: my first version let Gemma judge my answers. The 1B model was far too agreeable. I told it I'd solve a bracket-matching problem by reversing a linked list, and it called that a great approach.

So I split the jobs. Each problem has a few keywords for its key idea (for Valid Parentheses, the word "stack"). The code checks my answer for them and decides the verdict. Gemma only writes the feedback, and it must follow the verdict. That made the small model reliable. The tradeoff is that a keyword check is simple. Saying the right word in a wrong solution can still pass.

Other things I fixed while testing: it cut me off during long answers, so I raised the pause time and the answer limit.

Why Does Open Innovation Matter?

It works offline. After the one-time model downloads, it needs no internet, which matters on a trail with no signal.
My voice stays on my laptop. Nothing is sent to a server I don't control.
It costs nothing to run. No API key and no usage bill.
I could change how it behaves. I rewrote the coach's rules and added the code-checked verdict myself. With a closed chat API, I wouldn't have been able to work around the model's weaknesses this way.

Prize Categories

Best Use of Gemma

Comment Shield: read your YouTube comments without the sting

Marturu Ramyasri — Sat, 03 Oct 2026 19:27:47 +0000

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

What I Built

Comment Shield is a small website for YouTube creators. You paste a video link, and it sorts the comments into Positive, Negative and Neutral.

I built it for my friend Manasa, a YouTuber. Scrolling through hundreds of comments looking for feedback means constantly running into the mean ones, and that wears her down. Creators need the feedback, but they shouldn't have to take a hit every time they check.

Who It's For and Why

Manasa my friend makes videos on Black Pink Edits. She told me whenever she sees a few comments on a single category(negative/positive) she just assumes that this is the overall review of that video. But when we differentiate the comments into different categories like positive or negative or neutral, it lets her know the actual truth of the opinion on the video. So I didn't just want a sentiment counter. I wanted something that changes how she reads comments.

Features

Positive first: the app opens on the kind comments.
Gentle mode (on by default): negative comments are blurred, and she taps one only when she's ready to read it.
🚩 Harsh flag: strongly abusive comments are marked so she can skip them completely.
Summary bar: shows the positive / neutral / negative split, so she can see that most viewers are on her side.
Private by design: no backend and no accounts. The API key and comments stay in the browser.

Demo

ramya2006sri.github.io

Code

ramya2006sri / comment-shield

🛡️ Comment Shield

Comment Shield is a privacy-first web application designed for YouTube content creators. It automatically analyzes and categorizes comments on your videos into Positive, Neutral, and Negative sentiments—allowing you to read kind words first and engage with feedback on your own terms.

✨ Features

Positive First: Loads encouraging and supportive comments first by default.
Gentle Mode: Blurs negative comments automatically. Tap any blurred comment to reveal its text when you're ready to read it.
🚩 Harsh Flagging: Highlights strongly abusive or harsh comments so you can easily identify or skip them.
In-Browser Sentiment Analysis: Uses an open-source AI model running directly in your browser. Your comments never leave your device.
Summary Breakdown: Visual sentiment distribution bar showing the percentage split of positive, neutral, and negative reactions.
Fallback Support: Includes a lightweight local lexicon/word-list engine as a backup if the AI model fails to load.

🛠️ How

…

View on GitHub

(https://dev-to-uploads.s3.us-east-2.amazonaws.com/uploads/articles/zs541t1itcvrsegcxhwj.png)

How I Built It

Stack: one HTML file, vanilla JavaScript, no framework and no build step.
Data: YouTube Data API v3 (commentThreads), up to 500 top-level comments per video.
Open-source AI: the open-weight model twitter-roberta-base-sentiment-latest (RoBERTa trained on social media text), run locally in the browser with Transformers.js.
Fallback: a small word list is used if the model can't load, and it also helps flag very harsh comments.
Hosting: GitHub Pages.

Limitations: it can misread sarcasm, it works best in English, and the first load downloads the model (cached afterwards).

What Manasa Said

"It helped me know how many of the audience are in favour of me and also the different categories of comments helped me a lot. Instead of manually checking all the comments this platform helped me to reduce time consumption and also great project."

Why Does Open Innovation Matter?

A closed sentiment API would need every creator's comments sent to a third party and billed per call. With an open-weight model running in the browser, the comments stay private, it's free to use, anyone can read the code to see how comments are judged, and anyone can fork it and swap in a model for their own language.

DEV Community: Marturu Ramyasri

Walk Tutor: a voice DSA coach that runs offline on my laptop

What I Built

Code

ramya2006sri / walk-tutor

Walk Tutor

How it works

Setup

How I Built It

Why Does Open Innovation Matter?

Prize Categories

Comment Shield: read your YouTube comments without the sting

What I Built

Who It's For and Why

Features

Demo

Code

ramya2006sri / comment-shield

🛡️ Comment Shield

✨ Features

🛠️ How

How I Built It

What Manasa Said

Why Does Open Innovation Matter?