AI Personal Assistant
• Built an RAG-based AI Companion using LLaMa for responses and ChromaDB as vector database, Live 2D models
for character, and with BERT based sentiment analysis for interactive facial expressions.
• Increased relevance of responses by 50% while maintaining various story lines based on a predefined character outline
passed to the model by a PDF document.
• Added Playwright as an MCP tool to access web browser content, enabling real-time content retrieval from multiple
web sources
SaleriOS
• Engineered a real-time Operating System (RTOS) based on microkernel architecture for the Raspberry Pi 4, leveraging QEMU for efficient development, testing, and debugging in a virtualized environment.
• Implemented foundational OS components, including a custom C standard library, boot mechanism for system initialization, and a TTY console for essential command-line interaction and debugging.
• Developed a complete, self-contained software stack from scratch for bare-metal hardware, with memory and process
management.
ExploreSat
Developed a multi-spectral land-cover segmentation pipeline using six-band Sentinel-2 data, Google Earth Engine, Planetary Computer, SegFormer, and DeepLabV3+, and deployed a FastAPI service for batch inference and GeoTIFF export.
CrowdSentinel
Led development of an IoT crowd-risk system streaming edge video from AMB82 via Flask to AWS Kinesis Video Streams, with YOLO26 and LSTM models for real-time stampede-risk classification.
Sheild
Created a Zeek-based intrusion detection pipeline using flow features and TLS JA3 fingerprints, with a hybrid Transformer autoencoder, Random Forest, and Isolation Forest detection system and weighted risk scoring for automated threat mitigation.
