Abhinav Sachdeva

• Researched and designed a retrieval model based on transformer architecture, leveraging Python & Pytorch that can predict foot pressure from image & 3D models for physical activity such as Tai-chi with an average relative error of 5%.
• Developed a Variational Autoencoder (VAE) end-to-end using PyTorch, trained on 1.36M foot pressure samples, achieving 99.9% reconstruction accuracy.
• Trained, validated and tested the designed generative models in a remote Linux environment, leveraging anaconda to setup isolated, reusable environments.
• Improved an AI application using Python and Streamlit to generate SMPL(3D) body models using custom data input, enabling use-cases beyond original implementation.

• Graded 85 students for a graduate level course and mentored 140+ students in Python, improving course retention by 20% through office hours & code reviews.
• Prepared 100+ coding exercises and graded 200+ assignments in a Python based Object Oriented Programming course.

• Developed a full-stack job portal using Jenkins for CI/CD, serving front-end through Nginx, leveraging Docker and Kubernetes(K8s) for container orchestration, improving deployment time by 50% & deployed on Azure AKS.

Backend(built twice using Java, Spring Boot & Node, Express)

• Implemented a RESTful backend with Express.js, Node.js & Typescript with JWT authentication and role-based access, leveraging PostgreSQL & Prisma for database, hosting it on Azure DB.
• Implemented RESTful backend with Java & Spring Boot, with JWT authentication and role-based access, leveraging leveraging PostgreSQL & Hibernate/JPA for database, hosting it on Azure DB.

Frontend

• Designed the front-end in React, Next.js, JavaScript, HTML, & CSS, in under 10 hours.

• Deployed a donation seeking platform, a responsive web app built using HTML, TypeScript, jQuery, CSS.
• Leveraged Docker to containerize the website, deploying on Azure. Using Stripe API to facilitate donations, raising $1000 for the cause.

• Delivered a full-stack F1 fantasy app featuring a custom Java & Spring Boot backend that translated complex race results and betting rules into real-time fantasy points.
• Built the database using PostgreSQL, using Hibernate/JPA for management, hosted on AWS RDS
• Engineered a dynamic user interface with React and TypeScript, featuring multiple interactive elements.
• Established an automated CI/CD pipeline with Jenkins, containerized with Docker & deployed using AWS EC2.

• Engineered a full-stack job application chrome extension for 30+ beta users, automating outreach & tracking.
• Implemented a robust Gmail API integration using OAuth 2.0, enabling a 1-click email feature that users leveraged to generate over 1,500 email drafts.

Backend (built twice using Java, Spring Boot & Typescript, Node, Express)

• Built and deployed the backend infrastructure on AWS EC2 using Docker, featuring a RESTful API with Node.js, Express, secure JWT authorization, and a PostgreSQL database managed by Prisma ORM.
• Built and deployed the backend infrastructure on AWS EC2 using Docker, featuring a RESTful API with Java, Spring Boot, secure JWT authorization, and a PostgreSQL database managed by Hibernate/JPA hosted on AWS RDS.

• Created Spotified, an open-source Spotify Web API SDK for Browser and Node, earning 100+ GitHub stars.
• Created Spotified, an open-source Spotify Web API SDK for Browser and Node, with 1k+ monthly NPM downloads, adopted by 100+ developers.
• Delivered 98% test coverage with Jest for high reliability across ESM and CommonJS builds.
View on Github

• Built a Chrome Extension enabling users to group links into tab groups via right-click, reducing steps by 60% and eliminating manual tab dragging.
Applied core UI/UX principles to design intuitive interactions, Implemented dynamic context menus, debounced UI updates, and error handling for seamless User Experience
View on Github

• Developed a watchOS app to estimate ball throw speed and spin using inertial sensor data.
• Implemented a speed estimation algorithm with a median error of 1.4 m/s.
• Designed a pipeline with server-side processing using Flask and SciPy, and real-time visualization with Chart.js.

• Engineered an Image Caption Generator model, based on the ”Show, Attend and Tell” paper, utilizing TensorFlow and Keras to experiment with different CNN's (image feature extractors) models.
• Achieved a 1.87% increase in BLEU scores after 20 epochs of training on 30k images, with potential for further improvements through more epochs and a larger dataset.

• Developed a custom dynamic storage allocator for C programs, implementing `malloc`, `free`, and `realloc` functions.
• Designed and optimized heap consistency checker to ensure robustness and debug allocation issues effectively.
• Achieved efficient memory management with 16-byte alignment and minimized fragmentation through advanced policies.

• Developed a P2P file sharing system with a central server for managing file chunks and peers.
• Implemented parallel downloading by dividing files into chunks and retrieving them from multiple peers simultaneously.
• Validated file integrity using hash-based verification to discard corrupted chunks.
• Designed a multithreaded system using Python's socket programming and threading modules.
• Provided features for file registration, list retrieval, and real-time download progress tracking.

• Trained and fine-tuned GPT model on 5,500 lines of Robert Frost's poetry to generate stylistically coherent text leveraging PyTorch and Jupyter Notebook.

• Built a movie recommendation engine leveraging collaborative filtering (matrix factorization), trained on 32M user ratings from MovieLens. Deployed via Flask for scalable API access, demonstrating end-to-end ML systems design and large-scale data mining