Projects

Swarm Builder: An multi-agent system to build new systems.

Developed an AI-powered framework that dynamically generates OpenAI Swarms based on user prompts.
Enables rapid prototyping of multi-agent systems, reducing setup time from hours to minutes.
Includes a Streamlit interface supporting text and audio input for seamless swarm creation.
Prebuilt examples include Bitcoin analysis, Twitch streamer tracking, Twitter discussion summarization, and more.
Automates 90% of the swarm setup, requiring minimal user customization.

Developed a real-time multiplayer drawing game using WebSockets, OpenCV, and Python.
Implemented webcam-based hand tracking for gesture-controlled drawing and power-ups.
Designed interactive gameplay with a shared canvas, unique power-ups, and score tracking.
Enhances virtual engagement by enabling creative, gesture-based multiplayer interaction.

Built a RAG based search tool with ChromaDB that retrieves and summarizes game solutions from the internet in real-time (YouTube, Reddit, and web guides).
Extracts key timestamps from YouTube videos and provides direct sources when scraping is restricted.
Helps gamers save time by instantly finding the best strategies, walkthroughs, and hidden details without endless searching.

Built and deployed a web application that visualizes Neural Style Transfer (NST) in real-time.
Enhanced interpretability by integrating explainability techniques (XAI) like Feature maps, Gram-matrix and Grad-CAM visualization.

Completed data collection process from IRB approval to open-source release on HuggingFace.
Collected over 500 audio samples along with demographic metadata.
Ensured ethical standards with anonymization and open-sourced the dataset to promote inclusivity.

Developed an unsupervised CNN-based segmentation algorithm.
Achieved 92.54 Dice coefficient accuracy in segmenting peck-damaged rice grains.
Earned the highest grade for conducting research in a novel field.

Designed a miniature rolling capsule with a CCD camera for pipe inspection.
Developed a U-net model for detecting defects on steel surfaces with high accuracy.
Achieved a 2% improvement over baseline results on steel defect detection dataset.

Predicted Remaining Useful Life (RUL) in a fleet of aircraft engines with multiple failure modes.
Used ragged tensors with CNN and RNN-based models for time series prediction.
Learned LSTMs, GRUs, Bi-directional LSTMs, and Attention models for time series data.

Developed the backend of a customizable CNN architecture for traffic sign recognition.
Built baseline model for GTSRB dataset with 43 traffic sign classes.
Learned data augmentation, CNN layers, hyperparameters, and UI design.

Extracted facial landmarks and temporal signals to differentiate between authentic and spoof videos.
Used OpenFace library and learned basics of MATLAB.
Worked with non-contact face video data for heart rate estimation.

Developed a scheduling system for a simulated microgrid controller.
Built a multivariate time-series forecasting model for renewable energy prediction.
Gained experience with RNNs and convex optimization algorithms.