Projects

Engineered a closed-loop Brain-Computer Interface (BCI) using a Muse 2 gateway to stream real-time EEG/PPG data, training a 1D-CNN and LSTM pipeline to predict cognitive fatigue with a 30-minute forecasting window. Architected a Mixture-of-Agents (MoA) intervention system in Node.js and LangGraph via the Featherless.ai API, dynamically routing between 3 LLMs (DeepSeek, Qwen, KimiK2) to autonomously block distracting applications.

<> Next.js, Node.js, React, Flask, PyTorch, LangChain, LangGraph, Featherless.ai

Sprinklify is an intelligent irrigation assistant that integrates NLP, computer vision, optimization, ML, and IoT into one seamless platform. Users can simply describe their lawn (“7m by 5m lawn with rounded edges…”) and Gemini parses the text into structured GeoJSON, validated by spaCy + regex. Or they can upload a photo or drone video, and our Flask CV pipeline (OpenCV + Shapely) segments lawn areas and sprinkler points to auto-generate GeoJSON. Once a lawn is defined, our solver + reinforcement learning agent generates per-sprinkler rotation and flow schedules that minimize overspray and water use while maximizing coverage. Gemini also powers three LLM “agents” for users. We also had real hardware simulation with Arduino-based IoT sprinklers in a physical testbed!

<> React, Tailwind CSS, Spring Boot (LLM/NLP/Auth0), Flask (CV/ML/Optimizer), Gemini API, spaCy, regex, PPO RL, OR-Tools, MATLAB, OpenCV, Shapely, MongoDB

Research: A novel variability-aware fuzzer that integrates principled dynamic variability-aware analysis within the fuzzing process to enhance configuration space exploration using new variability-aware seed selection and mutations to drive the fuzzing process.

<> Docker, C, C++, Python

Research: In this work, we identify critical gaps in current benchmarking infrastructure that prevent fair and systematic comparison of such fuzzers. We extend FuzzBench to support configuration-aware fuzzing.

<> Docker, C, C++, Python

An app to visualize gas pipeline data through an interactive dashboard for detection and prediction of hydrate formation using current gas injection metrics with a 95% accurate regression model for real-time predictions.

<> Flask, Next.js, ML