Hi, I'm Dhruv, a Computer Science and Electrical & Computer Engineering student at Cornell. I care about efficiency, in models, in code, and in the machines they run on. I enjoy working across robotics, AI/ML, and GPU/low-level programming. In my off hours, you'll find me watching and playing soccer, ping pong, pickleball, exploring new restaurants, and reading.
Cornell Tech, AI & Robotics LabML Researcher2026Extending inverse semantics (Tellex et al., RSS 2014) with a minimax objective for better human-robot communication. Evaluating VLMs for large emergency medicine video understanding systems.
InvestcorpSoftware Engineer2026Building LLM agents to automate & structure treasury inbox monitoring for downstream financial analysis.
Weill Cornell Meyer Cancer Center, CompBio LabML Researcher2024 – 2025Trained U-Net + Vision Transformer segmentation models for microscopy organoid segmentation.
Virginia Tech, Data Management LabSoftware Engineer2023Developed deep learning object detection models and a web application for scientists to automate nematode classification.
An email agent made to manage Gmail while keeping inbox data fully on device. It categorizes mail with Ollama models, answers questions about the inbox, and drafts replies that only send after approval.
(Python, Ollama, Next.js, Gmail API)
A control tower that keeps Claude Code sessions on budget and honest. It runs preflight token checks, watches for stalls, and writes landing reports that compare claimed changes against what actually ran.
(Node.js, CLI, lifecycle hooks)
A local dashboard for understanding group chat dynamics without messages ever leaving the Mac. It reads the iMessage database read only and computes over 15 metrics per person, from response speed to reaction networks.
(Python, pandas, SQLite, Streamlit)
A vision pipeline that recovers 3D soccer ball trajectories from flat broadcast footage. It calibrates the camera from pitch geometry, tracks the ball, then fits drag and Magnus force models to resolve depth.
(Python, computer vision, physics modeling)
A physics sim for testing whether an agent can learn to swing through a city on ropes. A 240 Hz solver models gravity, drag, and rope tension in an endless procedural skyline. An RL agent trained with the cross entropy method, built without any ML frameworks, flies it live below.
(JavaScript, canvas, physics, RL)
Swung live by an agent trained with RLexplore the sim ↗A live demo of machine learning happening right on this page. 80 tiny neural nets evolve to clear obstacles with no ML frameworks and no training data.
(TypeScript, canvas, neuroevolution)
An investigation into how reliably ML models diagnose brain tumors from MRI scans. It benchmarks CNNs, boosting, and multiplicative weight update ensembles, reaching 96% accuracy across datasets. Published in the Journal of Student Research.
(TensorFlow, Keras)
