All-Weather Multi-Spectral Satellite Fusion Network

PyTorch Multi-modal DL Patented

• Invented environment-conditioned cross-spectral attention — atmospheric state (clear/cloudy/smoky/night) directly modulates SWIR-thermal-visible band interactions
• Thermal super-resolution from 100m → 30m using learned Lagrange multipliers, preserving temperature accuracy within 0.2K
• 79% mAP vs. 67% single-network baseline (+18% relative gain)

Hardware-Adaptive Zero-DCE Edge System

PyTorch TensorRT NVIDIA Jetson Nano Patented

• Architected degradation-aware enhancement model superseding Zero-DCE; adaptive to high-variance noise and motion blur
• Deployed at 35 FPS / ~15ms latency on Jetson Nano via TensorRT INT8/FP16 quantization
• Dynamic hardware routing via gating mechanism along quality-vs-latency Pareto frontier

CNNs YOLO PyTorch Paper Submission in Progress

• 96% classification accuracy on MRI data via custom CNN architecture
• Integrated YOLO bounding-box detection for precise tumor localization
• Full ML pipeline: preprocessing → augmentation → training → Precision/Recall/F1/AUC evaluation

ArduPilot Pixhawk ROS2 C++

• GPS-denied navigation reducing flight path deviation by ~40% vs. open-loop baseline
• Coordinate frame transformations + IMU sensor fusion for stable dynamics in unstructured environments
• Designing decentralized multi-agent path planning for a swarm of 3 drones

VINS-Mono RTAB-Map ROS2 A* C++

• Engineered Visual Inertial Odometry (VIO) fusing monocular vision + high-frequency IMU for precise state estimation in GPS-denied zones
• Real-time loop closure and occupancy grid mapping via RTAB-Map
• A* path planning on live occupancy grid for dynamic obstacle avoidance

Next.js Express MongoDB TypeScript Gemini API

• Engineered a closed-loop active web vulnerability scanner featuring a real-time SSE attack console
• Deployed dynamic mitigation system with automated framework-specific patch generation (F → A+ grading)
• Designed responsive Bento dashboard with WebGL posture visualization and AI-driven threat briefings

PyTorch FAISS SQLite FTS5 NLP Transformers

• Fuses FAISS semantic vector search with SQLite FTS5 keyword query execution via reciprocal rank fusion (RRF)
• Implements custom PyTorch representation projection bridge mapping 768d BERT embeddings to LLaMA 4096d hidden space
• Quantized to 4-bit NF4 and optimized for local edge GPUs under 8GB VRAM limit using lazy-loading and offloading

Optimization Pathfinding Logistics Python

• Designed an intelligent real-time traffic and logistics optimizer for delivery route planning.
• Resolves vehicle routing problem (VRP) under time-window constraints and traffic congestion.
• Features dynamic re-routing with live traffic updates to minimize delays.