Built from every version of me I had to outgrow

youtube

leetcode

productHunt

devto

github

Shorya Sethia  • Peerlist

•	Implemented adversarial attacks using the Fast Gradient Sign Method on CIFAR10 to evaluate the proposed defence mechanism's effectiveness against both white-box attacks •	Designed a Defence pipeline involving Denoising Autoencoder and Block Switching Model •	Utilized Denoising Autoencoder to filter out perturbations from input images, enhancing the resilience of the model to adversarial noise. Trained it for multiple noise factors ranging from 0.01 to 0.25 for CIFAR10 •	Implemented Block-switching using 4 instances of modified VGG16. Re-Trained the combined model with parallel channels and a common upper model; parallel channels can disrupt gradient-based attack vectors. •	Successfully defended against strong FGSM attack and classification accuracy increased from 6.84% to 54.36%

Adversarial Attack and Defence

Implemented various (vanilla, deep convolutional, VAE, de-noising) autoencoders.

Autoencoders

- Made an Arduino based bot which takes the data of voltages given by solar panel in first lap while tracking a given 2D plane and once it is completed it decides the max voltage point and goes there in another lap. - Also solar panel is mounted on 2 servos which provides bi-directional (L-R and up-down) degree of freedom based on the input given by 4 LDRs.

Automated Solar Panel Positioning System

• Developed a production-ready blog platform with React, Redux Toolkit, and Appwrite Backend as a service featuring user authentication, protected slug-based routing, post management, and real-time content editing • Implemented TinyMCE rich text editor with Tailwind CSS for responsive design, role-based access control, SEO-optimized URL structure. Deployed on Vercel with automated CI/CD pipeline, performance optimization

Blogify - A full stack Blog Platform

An agentic approach for querying and analyzing UPI transaction data using natural language

ClrInsights

• Successfully generated MNIST-like images, achieving a Structural Similarity Index Measure (SSIM) of 61.88% • Modelled C-GAN by converting MNIST labels into dense Embeddings, enabling effective concatenation with input Gaussian Noise for the Generator and with real as well as reconstructed images for the Discriminator • Implemented and Deployed the model testing pipeline on Streamlit, leveraged the trained Generator to test and get performance of any model trained on MNIST, utilizing both real and generated images for evaluation

Conditional GAN based Model evaluator

Used MNIST on various ML and DL architecture like MLPs, VGGs, LeNet-5, CNN etc. Compared their accuracies and complexity.

Digit Recognition

• Architected an pipeline leveraging Microsoft's Table Transformer and XRef ID parsing to segregate textual data, tables, and visual content from annual reports, enabling automated document understanding • Developed robust support of 6 LLM clients to Generate Descriptions for both visual content (plots/images) and tables, concatenating these with text data to built knowledge graphs using LightRAG to provide depthful answers to queries, without relying on RAG-LLM, to enhance the overall relationships extraction process • Engineered an document transformation pipeline via Docling, Regex and Pandas to convert reports into web format with Downloadable Tables, reducing manual analysis time and streamlining information accessibility

FinChat: An Financial Document In-Depth Analy

• Secured 4th place in Qualcomm VisionX Challenge (prizes worth: ₹1.5 Lakhs) by developing a transformer-based reflection removal system with Cross-Scale Attention Module and Multi-Scale Fusion Module, achieving best-in-class PSNR of 23.61 and SSIM of 0.835 on Real dataset, with superior robustness against MSE attacks (PSNR: 20.15) and LPIPS attacks (PSNR: 22.02). • Engineered an adversarial-robust architecture leveraging transformer's global context modeling and cross-scale attention mechanism, demonstrating significant improvement over baseline WY19 model with SSIM scores maintained above 0.752 under MSE attacks and 0.768 under LPIPS attacks on Real dataset. • Designed a triple-loss optimization framework (adversarial discriminator + GAN + perceptual loss) with multi-scale feature processing capability, improving PSNR from 19.23 to 23.61 on Real dataset with adversarial training, while maintaining high SSIM scores across Nature, SIR², and Real datasets. • Implemented transformer-based feature extraction over conventional UNets/Autoencoders, achieving minimal PSNR degradation (only 2.11 drop under MSE attacks, 0.24 under LPIPS) on Real dataset compared to baseline degradation of 8.21 and 3.04 respectively. • Developed multi-scale input encoding pipeline with dynamic feature aggregation, demonstrating consistent performance across all three datasets (Nature: PSNR=20.97, SIR²: PSNR=23.02, Real: PSNR=23.61) while effectively handling both fine-detailed and blurred reflections.

Qualcomm VisionX: Reflection Removal

Regulatory intelligence that maps how rules evolve. Basically a tool to get structured reference extraction for SEBI circulars.

Relatio

• Engineered a TCP based bi-directional chat application in C++ using the Winsock API, achieving full-duplex communication. • Developed a custom, binary-safe file transfer protocol featuring buffered transmission for efficient data transfer.

Socket Programming

Fine tuned InceptionV3 for cat-dog classification. Achieved 95.75% test accuracy.

Transfer Learning (InceptionV3) for Cats - Do

Guide: Prof. Dr. Nirmal S. Punjabi from Koita Centre for Digital Health, IIT Bombay • Modelled Respiratory Rate Prediction model by engineering temporal features from ECG & PPG signals; achieved 90.44% R² score using Random Forest, outperforming XGBoost (81.78%) & LightGBM (56.98%) • Developed multi-class ECG arrhythmia classifier (5 categories) on MIT-BIH dataset, achieving best 97.981% accuracy through weighted ensemble of sequential RNN-based models LSTM (96.88%) and GRU (98.43%) • Optimized ensemble performance by implementing selective inference based on confidence threshold of 0.80

TrueMeasure: AI-Driven Insights from Wearable

• Built Whistler, an AI-native IIT placement intelligence platform serving 2,500+ users with 65K+ web interactions. • Developed an MCP (Model Context Protocol) server enabling AI agents to query whistler's data archive with authentication, rate limiting, and caching. MCP server has ~2.2k+ downloads. • Engineered a free-tier full-stack architecture using React, Express, MongoDB, Cloudinary, Vercel, and Render, including platform analytics, and rolling-window aggregation for scalable management.

Whistler

Pneumonia or Normal classifier based on Chest X-Ray using CNNs Was able to achieve 74.83% testing accuracy which is decent for medical tasks