CS & Math · Duke University
Developer. Founder. Researcher.
Building and shipping things that matter.
Hi, I'm Siddhant, a Computer Science and Mathematics student at Duke. I love building things at the cusp of research and real products - stuff that's interesting to make and actually matters once it's out. I'm grateful to say that some of what I've shipped is already making an impact. Outside of coding and research, you'll find me at the gym, playing cricket and FIFA. I'm always down to meet interesting people, so feel free to reach out!
Solving student doubts, managing a help room, and helping students achieve good grades.
Investigating optimal group granularity in RLHF via GRPO on Gemma-2B, benchmarking across 106 countries to reduce alignment bias by 30%.
Conducting due diligence, market research, and startup evaluation for potential investments, and working on UI/UX feedback for STAND, an SF-based startup.
Building ML models for computer vision and finger recognition using MediaPipe, contributing to research on gesture-based human-computer interaction.
- Full-stack NFL player web-based equity marketplace (back-d.com) where fans buy and sell athlete shares like stocks using an AMM pricing engine with a pre-defined platform fee and Stripe-powered deposits
- Built with Next.js 14, Node.js, PostgreSQL, and Prisma ORM, featuring Clerk auth, live price charts, and fan portfolio dashboards
- Built a custom CUDA transformer attention kernel from scratch, achieving 15.6x faster inference than baseline and 1.9x faster than PyTorch on Duke HPC (RTX 2080 Ti) via memory coalescing and shared memory tiling
- Diagnosed less than 1% GPU utilization through CUDA event profiling and optimized memory access patterns, reducing latency from 1.88ms to 0.069ms at seq=512
- AI-powered Chrome extension that scores every Gmail sender's trustworthiness from 0-100 in real time, detecting phishing attempts, spoofed identities, and urgency manipulation tactics used by bad actors
- Built with Manifest V3 and Vanilla JS on the frontend, backed by a Python FastAPI server handling WHOIS domain lookups, rule-based NLP analysis, and sender reputation scoring
- Owner of a fully monetized 4-year-old tech YouTube channel centered around CS innovations, real-life problem solving, and programming tutorials in Python, Java, automation, and AI tools - with a global viewer base spanning the USA, Canada, India, and beyond
- Grown to 1,200+ subscribers and 500,000+ channel views, with standout projects including a viral Gmail storage hack series (200,000+ views), a Python Personal Assistant, and a NewsBot - all built to solve real problems with code
- Part of the mechanical team building and programming an industrial-size robot - designed a custom MK4i L2 swerve drive chassis and 2 telescopic arm contraptions
- Won the Industrial Design Award sponsored by General Motors for integrating 8 different parts in the shooter mechanism
- Developed a smart IoT-based autonomous plant watering system using NodeMCU ESP8266, soil moisture sensors, and solar energy
- Served as Head of Programming - wire planning, LCD soldering, and full Arduino system programming
- Engineered a robotic arm prototype mapped to EMG sensors on a user's arm using ESP32, servo motors, and Arduino
- Submitted to Plaksha University Electrical Engineering Department for demonstration to undergraduate students
- Authored a research paper under the supervision of Dr. Martin Sewell (University of Cambridge), predicting Indian gold prices using LSTM neural networks trained on 24 years of data with 96.79% accuracy
- Built using Python, sklearn, Keras, and TensorFlow - model runs over 1,000 epochs to maximize prediction accuracy
- Devised a methodology to legally optimize Gmail storage using POP/IMAP email transfer via Gmail Shuttle Cloud Migration
- Video series garnered over 200,000 views and 500+ positive comments
- Built a fully operational Personal Assistant using pyttsx3, webbrowser, wikipedia, datetime, and speechrecognition libraries
- Garnered over 5,000 views on YouTube with consistently positive feedback