Freelancers using PyTorch in MumbaiFreelancers using PyTorch in Mumbai
ProjectsPeople
Anurag Nagare's avatar
Anurag Nagare
Mumbai, India
I’m an AI & Machine Learning engineer with expertise in deve
I’m an AI & Machine Learning engineer with expertise in deve
Cover image for What your attention heatmap isn't
What your attention heatmap isn't telling you Everyone's staring at attention heatmaps and calling it "interpretability." Almost nobody's asking whether a single attention map actually tells you what the model used to make its decision. It doesn't. Not on its own. A raw attention map from layer 8 shows you what layer 8 attended to. It says nothing about how that signal got mixed, diluted, or overwritten by every layer before and after it. Attention rollout fixes this — and I built a walkthrough to show why it matters. Here's what makes it more than a "pretty heatmap" demo: Instead of visualizing one layer's attention, I traced how information actually flows through the full transformer stack. → Every layer's attention matrix is extracted, per head, per token → Multi-head attention is averaged, then combined with the residual connection (identity + attention) — this is the step most tutorials skip, and it's the one that actually matters → The combined matrices are matrix-multiplied layer by layer, rolling attention forward from input to output → The result: a single map showing genuine token-to-token influence across the entire network, not just one layer's snapshot The overlay shows you everything: → Per-layer attention vs. rolled-out attention, side by side → Token importance scores overlaid directly on the input text → A comparison view: which tokens raw attention says "matter" vs. which ones rollout says actually matter → Head-level breakdown so you can see which heads specialize vs. which are noise No black box. No "trust me, the model looked here." Just linear algebra, applied honestly across every layer instead of cherry-picking one. Built with PyTorch + HuggingFace Transformers + Matplotlib. Runs on any pretrained transformer, fully offline. ⚠️ Important: attention rollout is an approximation, not ground truth. It assumes attention is the primary information pathway, which ignores MLP layers and can still mislead for very deep models. Treat it as a debugging lens, not proof of causality.
0
74
Cover image for Everyone's racing to add biometrics
Everyone's racing to add biometrics to logins. Almost nobody's asking what happens when you can't — or shouldn't — touch the sensor. Shared kiosks, clinical settings, accessibility needs, hygiene-sensitive environments. Fingerprint readers and face unlock assume contact or a stored faceprint. Sometimes you want authentication that touches nothing and stores no biometric image of you at all. So I built GestureAuth — a contactless authentication system where your "password" is a sequence of hand gestures performed in front of a standard webcam.
1
68
Cover image for Everyone's talking about quantum computing.
Everyone's talking about quantum computing. Nobody's using it to feed farmers. India loses 20–30% of its crop yield every year to diseases and pests. Not because farmers don't care — but because early detection is hard, expensive, and inaccessible to the people who need it most. The existing solutions? Either a basic image classifier trained on lab-perfect photos that fail in real field conditions, or an agronomist visit that costs time and money most small farmers don't have. So I built QuantumEdge AgriGuard — a hybrid Quantum Neural Network app where a farmer can photograph a diseased leaf on their phone and get an instant diagnosis in under 5 seconds. Here's what makes it different from just another plant disease detector: Instead of a pure classical CNN, I built a hybrid architecture — a ResNet/EfficientNet backbone extracts visual features, then passes them into a Variational Quantum Circuit (VQC) for the final classification. The quantum layer uses angle embedding + StronglyEntanglingLayers, which gives it a measurable edge on small, noisy datasets — exactly the kind of data you get from Indian field conditions. The app doesn't just tell you what disease it is. It gives you: → Confidence score → Organic + chemical remedies (India-specific) → Yield impact estimate → A live classical vs quantum accuracy comparison so you can see the difference yourself I tested the quantum advantage claim honestly — ran both models on the same downsampled PlantVillage dataset and tracked accuracy, F1-score, and inference time side by side. The results are on the dashboard. No hand-waving. Built with PennyLane + PyTorch + Plotly Dash. Designed to run on simulators today and on QpiAI-Indus 25-qubit hardware tomorrow.
1
0
80
Cover image for A Neural Network Visualization
A Neural Network Visualization Tool that demystifies AI! I built an interactive web application using Flask and PyTorch that doesn't just recognize handwritten digits it shows you exactly how the AI "thinks." When you draw a digit (0-9) on the canvas, the app processes it through a Convolutional Neural Network and generates a real-time visualization of every layer: from edge detection in the first convolutional layer, through pattern recognition, pooling, and feature extraction, all the way to the final classification. The tech stack includes Python, Flask, PyTorch, and vanilla JavaScript for the frontend. What makes this unique is the educational aspect each layer's activations are visualized using matplotlib, showing the 32 filters in Conv Layer 1, the 64 filters in Conv Layer 2, and the 128-neuron fully connected layer.
1
75
PyTorch(4)
Chirag Patankar's avatar
Chirag Patankar
Thane, India
Freelance Web Developer & AI Enthusiast helping brands build
Freelance Web Developer & AI Enthusiast helping brands build
Cover image for Weld Defect Classification System
Weld Defect Classification System
0
9
Cover image for Toxic Comment Classifier and Moderator App
Toxic Comment Classifier and Moderator App
0
7
Cover image for AI Customer Support Bot - MCP Server Development
AI Customer Support Bot - MCP Server Development
0
12
Placeholder project card media
View more →
PyTorch(1)
Mayur Parab's avatar
Mayur Parab
Mumbai, India
A data scientist with strong MLOps expertise.
A data scientist with strong MLOps expertise.
Cover image for Image Segmentation with Meta's SAM2 Model.
Image Segmentation with Meta's SAM2 Model.
0
7
Cover image for End-to-End ML Pipeline Automation
End-to-End ML Pipeline Automation
0
6
Cover image for End-to-End Data Pipelines.
End-to-End Data Pipelines.
0
5
Placeholder project card media
View more →
PyTorch(1)
Akshar Samudrala's avatar
Akshar Samudrala
Mumbai, India
Data Scientist who can do Project Management and animation
Data Scientist who can do Project Management and animation
Cover image for Text Simulation Game using NLP
Text Simulation Game using NLP
0
5
Cover image for Requirements Traceability Analysis using NLP
Requirements Traceability Analysis using NLP
0
3
Cover image for Speech Emotion Recognition System Development
Speech Emotion Recognition System Development
0
4
Placeholder project card media
View more →
PyTorch(3)
Akshat Singh's avatar
Akshat Singh
Mumbai, India
Full-stack expertise. AI integration.
Full-stack expertise. AI integration.
Cover image for LLM Agents for Simple Task automation
LLM Agents for Simple Task automation
0
15
Cover image for Dashboard and Parallel Services for a Gym
Dashboard and Parallel Services for a Gym
0
16
Cover image for Whatsapp and Discord Bot
Whatsapp and Discord Bot
0
23
Placeholder project card media
View more →
PyTorch(1)