Devarsh
Joshi

A self-hosted feature flag platform that changes what an app does in under a second, without a deploy. Flags support targeting rules and sticky percentage rollouts, and each saved change is published over Redis so every evaluation pod pushes it to connected apps through WebSockets. It includes a React dashboard with change review and protected environments, scheduled release plans, a TypeScript SDK with React hooks, and Kubernetes and Azure deployments with a preview environment per pull request.

• 0.3 to 0.7sClick to live update
• 500Automated tests
• 95%+Engine coverage

• .NET 10
• React
• TypeScript
• SQL Server
• Redis
• Kubernetes
• Azure

A web application that turns an ordinary video into an animated 3D Gaussian splat you can orbit and scrub through in the browser. It recovers the camera trajectory to classify each clip, then either fuses orbiting footage into a single 3D model or builds a spacetime model from a static background plus flow tracked gaussian trajectories. I wrote a differentiable splat renderer in PyTorch that optimizes every gaussian by gradient descent on Apple silicon, a custom binary scene format, and a WebGL viewer that interpolates motion continuously on the GPU.

• 300kGaussians optimized
• 2 to 10×Smaller scene files
• Apple MPSNo CUDA required

• PyTorch
• WebGL
• Three.js
• Next.js
• FastAPI
• OpenCV
• TypeScript

A deep learning model that detects lung cancer and classifies pneumonia from chest radiographs. I implemented Grad-CAM heatmaps to make each prediction interpretable, and applied test time augmentation to improve recall on the underrepresented class.

• 91.03%Accuracy
• 98.46%Recall
• Grad-CAMExplainability

• PyTorch
• Computer Vision
• CNNs
• NumPy
• Grad-CAM

A study platform that converts academic PDFs into short vertical videos using AI generated voiceovers and summaries. I built the natural language processing pipeline and contributed to the React front end. The platform was evaluated with a group of over 20 students.

• 200+Documents processed
• 60+ hrsSaved per month
• 40%Better recall

• ReactJS
• Flask
• OpenAI
• PostgreSQL
• Kotlin
• NLP

A mobile fitness application for logging workouts, tracking streaks and visualizing progress. I implemented the unit testing suite and continuous integration pipeline, and integrated media uploads through the Cloudinary API.

Repository private. Code walkthrough available on request.

• 50+Users
• 98%Bug-free deploys
• 300+Media uploads

• Python
• KivyMD
• FastAPI
• Matplotlib
• CI/CD

An exploratory analysis of over 10,000 drug reviews, presented as an interactive Streamlit application with visual reports intended to support public health research.

• 10,000+Records analysed
• 15+Visual reports

• Python
• Pandas
• Seaborn
• Streamlit