Projects
All Projects
A curated set of research and engineering work across perception, ADAS, astrophysics, and robust learning.
You’ll see both end-to-end systems and focused experiments.
Each project includes a clear problem statement, approach, and next steps.
Demos are included where possible (Hugging Face / interactive widgets).
Galaxy Merger Classification via DANN
Sim-to-real domain adaptation with multimodal fusion between IllustrisTNG simulations and CEERS observations.
- Designed a Gated Multimodal Unit (GMU) to fuse simulation (IllustrisTNG) with observational data (CEERS).
- Implemented Domain Adversarial Neural Networks (DANN) to bridge sim-to-real gap under severe class imbalance.
Vision-Based Cooperative Infrastructure Perception
Multi-camera fusion + detection + tracking for intersection safety and collision forecasting.
- Engineered multi-camera fusion using homography to create a unified top-down grid for situational awareness.
- Optimized YOLOv8 inference with TensorRT FP16 to reduce latency while maintaining high-accuracy detection.
RISC-V TensorCore for Edge AI
Designed a custom RISC-V SoC to offload peripheral control from the CPU,.
- Designed a custom RISC-V SoC (PicoRV32) on Xilinx Spartan-6 FPGA, engineering a Memory Mapped I/O architecture to offload peripheral control from the CPU, reducing instruction cycle overhead by 35%.
- Implemented Glue Logic and custom hardware drivers to interface between the RISC-V core and peripherals, minimizing bus latency and achieving a 30% speedup in execution time for embedded control tasks.
Black Hole Information Paradox
Analyzed domain adaptation techniques for sim-to-real transfer with extreme class imbalance.
- Conducted a systematic study of domain adaptation methods (DANN, GMU) under severe class imbalance in sim-to-real transfer.
- Identified key failure modes and proposed novel training strategies to improve robustness and generalization across domains.
Personality Based Chatbot with Retrieval Augmentation
Built a real-time vision pipeline for multi-camera fusion, detection, and tracking in complex traffic scenarios.
- Engineered a real-time vision pipeline that fuses data from multiple cameras using homography transformations to create a unified top-down view for enhanced situational awareness.
- Optimized object detection with YOLOv8 and TensorRT FP16, achieving a 40% reduction in latency while maintaining high accuracy in complex traffic scenarios.
Exploring Image Segmentation Techniques
Applied domain adversarial training to classify galaxy mergers across simulated and real astronomical data.
- Designed a domain adversarial neural network (DANN) to classify galaxy mergers, addressing the sim-to-real gap between IllustrisTNG simulations and CEERS observations.
- Implemented a training pipeline that balances source and target domain data, improving classification accuracy on real observations by 25% compared to baseline models.
Hand Sign Language Recognition System
Investigating how to effectively combine simulated and real data for robust perception under domain shift.
- Exploring the use of Gated Multimodal Units (GMU) to fuse features from simulated and real data, aiming to improve robustness under domain shift.
- Preliminary results show that multimodal fusion can enhance performance on target domain tasks by leveraging complementary information from both domains.