Vedant Sheel

A handwritten-digit classifier that runs entirely in the browser. Draw a number, hit run, and watch the whole network light up in 3D as it thinks. Every neuron, every weight, and every number is real, nothing is faked or pre-rendered.

• •Trained a 784 → 64 → 48 → 32 → 16 → 10 network in PyTorch to 96.9% test accuracy, then hand-wrote the forward pass in TypeScript so inference runs in the browser with no ML runtime behind it.
• •Rendered the network as a live Three.js scene driven straight off that forward pass. Layers of neurons are wired by their real weights, animate as activations propagate, and stay clickable down to the individual number.

PyTorch · TypeScript · Three.js · React Three Fiber

An autonomous drone system that picks a building or property, maps it in 3D, and flies itself to shoot cinematic footage, no piloting required.

• •Built an autonomous drone system that identifies a user-selected building or property, constructs a 3D representation of the environment, and detects visually significant viewpoints to automatically capture cinematic aerial footage without manual piloting.
• •Implemented optimization-based mission planning using viewpoint scoring, trajectory generation, spline smoothing, and OR-Tools routing to produce safe, physically feasible, and visually optimized flight paths.

Computer Vision · 3D Reconstruction · Path Planning · OR-Tools

A Raspberry Pi robot running QNX OS that understands plain-English objectives and figures out how to execute them itself, with no hand-written computer vision or control logic.

• •Built a modular, MCP-powered (Model Context Protocol) system that lets a Raspberry Pi robot running QNX OS understand and execute natural language objectives through real-time reasoning and dynamic hardware control.
• •Connected a Flask server exposing motion and vision endpoints to an MCP server, letting the LLM plan, act, capture images, and self-correct without traditional computer vision or manual logic.

MCP · QNX OS · Flask · Raspberry Pi · Gemini · Claude

A multi-model system that predicts how pulmonary fibrosis will progress from CT scans, accurate to within ±47 mL of lung capacity, now being validated on real clinical data.

• •Developed a first-of-its-kind multi-model system combining a self-attention CNN with regression to predict pulmonary fibrosis progression, reaching a ±47 mL absolute error rate on lung capacity decline.
• •Integrated a watershed segmentation algorithm for lung volume estimation and Hounsfield Unit normalization for CT preprocessing, outperforming traditional methods.
• •Working with the Firestone Institute for Respiratory Health to validate the model on clinical data, aiming toward peer-reviewed publication.

Self-Attention CNN · Regression · TensorFlow · OpenCV

An autonomous rover that patrols forests looking for wildlife, fires, and waste, cutting manual monitoring time by 70%.

• •Developed an autonomous environmental monitoring rover capable of detecting wildlife, fires, waste, and plant life with 94% classification accuracy using YOLOv8.
• •Integrated high-precision GPS mapping with a MongoDB-backed data pipeline, reducing manual forest monitoring time by 70% compared to traditional field surveys.

YOLOv8 · GPS Mapping · MongoDB