|

This is a research project focused on using deep learning to improve the early detection of gastrointestinal (GI) cancers. We’re developing a segmentation model that can accurately identify and outline abnormal regions in medical images — helping doctors spot signs of cancer at a much earlier stage. The goal is to make diagnosis faster, more consistent, and potentially life-saving. The work combines medical insights with AI to support real-world clinical needs, and it's currently under review for publication. Being part of this project has been both technically challenging and deeply meaningful — it’s one of the ways I’ve tried to apply AI for real impact in healthcare.

A real-time C++/CUDA engine that re-lights photos at an interactive 30 FPS, showcasing advanced GPU programming.

Saheli is a voice-based AI companion designed to feel less like a tool and more like a thoughtful friend. Inspired by the Hindi word for “female friend,” Saheli blends emotional intelligence with everyday support — helping users through calm conversations, mental wellness check-ins, and mindful productivity. Her voice is soft, grounded, and expressive — avoiding the robotic feel of most assistants. I’m building Saheli to work in Hinglish by default, with regional voice plugins planned. Technically, it’s powered by tools like Whisper for speech recognition and ElevenLabs for voice synthesis, with deployment possibilities across mobile, smart mirrors, browser extensions, or Raspberry Pi for edge-focused, privacy-first use. The goal is to create something that listens deeply, responds with care, and fits naturally into daily life. Saheli is still in the early stages, but it’s one of the most personal things I’ve ever built — and I’m excited to keep shaping it.

An event-driven replay and spectator modification for the game Mindustry, built with Java for enhanced game analysis.

AltSight Sentinel is a trading project where I explored unconventional sources of data to gain an edge in commodity markets. Instead of relying only on price and volume, the system combines satellite imagery and news sentiment to inform spread trades. I used OpenCV and AWS Rekognition to detect and count ships and storage tanks from satellite images, helping estimate supply flow. At the same time, FinBERT was used to analyze news articles for sentiment scores, which were processed through a Pandas pipeline. A custom portfolio optimizer then reallocated positions weekly using a CVaR (Conditional Value at Risk) constraint for better risk control. This project was my way of experimenting with creative alpha signals — the kind of thinking that's encouraged on real-world prop trading desks.