DEV Community: Shaikhhas

AI Fitness Coach: Real-time Exercise Form Analysis using Google AI Studio

Shaikhhas — Thu, 11 Sep 2025 09:16:02 +0000

This is a submission for the Google AI Studio Multimodal Challenge

What I Built

I built an AI Fitness Coach that analyzes exercise form from uploaded videos. This application helps users improve their squat and deadlift techniques by providing real-time feedback, identifying form issues, and suggesting corrective exercises.

Learning proper exercise form can be intimidating for beginners. Asking for help at the gym can feel uncomfortable, YouTube tutorials can be overwhelming with contradictory information, and personal trainers are often prohibitively expensive. The AI Fitness Coach aims to lower these barriers by providing an accessible, judgment-free way to check your form and get personalized feedback.

The AI Fitness Coach solves several problems:

Provides accessible form checks without requiring a human trainer
Helps prevent injuries by identifying risky movement patterns
Delivers personalized exercise recommendations based on specific form issues
Offers an educational tool for fitness enthusiasts to improve technique
Reduces the intimidation factor for beginners wanting to learn proper form

The GIF above demonstrates the key functionality of the AI Fitness Coach:

Uploading an exercise video
AI analysis of the movement
Receiving detailed form feedback
Reviewing specific issues with timestamps

How I Used Google AI Studio
I leveraged Google AI Studio's Gemini Pro Vision model to power the core functionality of the AI Fitness Coach. The implementation involved:

Video Analysis: Using Gemini Pro Vision to analyze exercise videos and evaluate form

Structured Response Generation: Configuring the model to return consistent JSON-formatted feedback

Prompt Engineering: Developing precise prompts that instruct the model to identify specific form issues in squat and deadlift exercises

Visual Understanding: Utilizing the model's ability to perceive movement patterns and body positioning across video frames

Multimodal Features

Form Analysis: Gemini identifies specific form issues like improper back angle, knee positioning, or weight distribution based on visual cues in the video.

Structured Feedback: The application parses the model's analysis into structured data that includes:

Overall form score
Positive aspects of the user's technique
Specific form issues with timestamps
Recommended exercises to address identified problems
Sources for recommendations

Technical Implementation

The application is built with:

Frontend: HTML, CSS, JavaScript
Backend: FastAPI (Python)
AI Integration: Google AI Studio's Gemini Pro Vision model
Deployment: Hosted on Vercel

The system processes video uploads, sends them to the Gemini API for analysis, and presents the results in an intuitive interface that allows users to review specific issues with synchronized video playback.

Important Disclaimer

This application is a proof of concept for demonstration purposes only. While the AI provides valuable feedback, it should be used with caution:

- The AI may occasionally misinterpret movements or provide incorrect advice
- Users should always prioritize their safety and comfort when exercising

The AI Fitness Coach demonstrates the potential of multimodal AI to make fitness guidance more accessible, but it's important to recognize its limitations and use it responsibly.

Contributors

This is a solo dev project by myself.

🔗 Project Repo

GitHub – GoogleAIStudioDemo

🔗 Project Link

Project

News Activity Map with Redis

Shaikhhas — Tue, 05 Aug 2025 09:01:21 +0000

This is a submission for the Redis AI Challenge: Real-Time AI Innovators.

What I Built

This project began as a simple RSS feed visualizer on a map. Over time, it evolved into a real-time political news intelligence platform powered by Redis, Telegram ingestion, NLP, and vector search.

The system is split into three main parts:

Telegram Scraper – collects news data, extracts named entities, and generates vector embeddings.
API – handles semantic search, clustering, timeline analysis, and pushes real-time updates via WebSockets.
UI – built with React + Vite + Leaflet to display the interactive map, article data, and activity timeline.

The (Very Rough) Dev Journey

Initial Proof of Concept

Scrape data from RSS feeds.
Normalize data for each source.
Run scraper on a schedule.
Use NLP (spaCy) for NER and geo-tagging (for mapping).
Generate article embeddings with all-MiniLM-L6-v2 for search and clustering.
Build a basic FastAPI backend for data retrieval.
Use React + Leaflet.js for the map UI.
Visualize activity with free country polygons and hover states.
Sidebar displays articles per country interaction.
24-hour cumulative article count shown as activity.
Store data in postgresql.

Second Phase

Switched to real-time ingestion via Telegram using Telethon (no more scheduling).
Normalized data once.
Upgraded NER using spacy/en_core_web_trf.
Retained embedding generation for clustering.
Started storing processed data in Redis.
Updated API to use Redis as the data source.

Third Phase

API now provides article counts grouped by hour to power a timeline.
Added NLP-based semantic search.
Search initially ran in-memory using the same embedding model (response time ~4–5 seconds).

Fourth Phase

Switched Redis storage to use HSET, ZADD, and byte arrays for vector search.
Created Redisearch index for fast KNN similarity lookup.
Improved search response time to ~0.5 seconds.
Switched activity metric from cumulative to time-weighted scoring.
Implemented similar article detection via embeddings.
Cleaned up and improved the UI.
Added map legend for better UX.

Fifth Phase

Final UI cleanup and layout polish.
Integrated Redis Pub/Sub for real-time notifications via WebSockets.
UI now receives live article alerts.
Refactored scraper and API codebases for maintainability.

Demo

Source Code

https://github.com/Shaikhhas/RedisDemo_PoC

Redis Usage

FT.SEARCH + KNN for vector-based semantic search and clustering.
ZADD and sorted sets for timeline activity.
Pub/Sub for pushing real-time notifications to the frontend.

Personal Wishlist / Future Ideas

Merge notification data into the main dataset

Precompute per-hour article counts for performance

Implement lazy loading or paging

Add dark mode

Add "Follow" functionality for developing stories

Expose a scraper ingestion API for other sources

This was a solo project built over about a week, based on a prior prototype.

Building a Global News Tracker with NLP and GeoTagging.

Shaikhhas — Tue, 15 Jul 2025 08:10:49 +0000

Prefacing this with - I'm not a writer. Why did I do this ? Because I wanted to see if I could do it with free resources and I was bored. A friend suggested posting here, so yeah...

Built a full-stack application that scrapes news articles from various RSS feeds and uses Hugging Face transformers for Named Entity Recognition (NER). Locations are refined via fuzzy matching with pycountry to automate geo-tagging, enabling display on a heat map built with React + Leaflet.

Designed as a proof-of-concept for real-time aggregation and visualization of global news using LLMs/AI.

Live Demo:

news-heat-map-demo-ui.vercel.app

Tech Stack:
• Frontend: React + Vite + Leaflet.js
• Backend: FastAPI + Uvicorn
• Database: Supabase (PostgreSQL)
• NLP: Python + HuggingFace Transformers
• Hosting: Render + Vercel

Key Features:
• Scheduled data scraping from RSS feeds
• Automated geo-tagging using NER + fuzzy matching
• Minimal, map-based UI for exploring story clusters

Everything after scraping is automated — from entity extraction and map display.

The country polygon are available freely which you can download, the RSS reader is a python script run through task scheduler when I log on. The UI could be improved further but I don't see anything I could add at the moment, the point was to give immediate feedback on what's and how much is happened where in the world.

NER picks up other stuff too which you could use for further filters, but that also requires cleaning.
Another addition would be grouping using embedding for related events.

The only manual part of the project right now is normalizing new RSS feeds.