Objective: Develop a Vision Language Action model for natural-language object manipulation in robotic environments

• Coupled a fine-tuned pi0-FAST action policy with a deterministic spatial grounder and symbolic planner
• Segmented LIBERO demonstrations into distinct subgoal phases using gripper open/close transitions
• Fine-tuned pi0-FAST with LoRA so the symbolic planner could direct the low-level policy through subgoals
• Achieved 92% and 82% success on Libero Spatial and Libero Long across 50 evaluation runs

Objective: Build a remote sensing multimodal chatbot for captioning, grounding, and VQA on satellite imagery

• Developed TRINETRA, orchestrating fine-tuned Qwen3VL, YOLO, SAM 2.1/3, and RemoteCLIP through a LLaMA 3.1 guide model
• Created GeoMeasure by combining fine-tuned SAM 2.1, SAM 3, and RemoteCLIP for robust numerical VQA
• Optimized a Tri-YOLO ensemble to normalize scale variance and support VLM grounding queries
• Surpassed baselines by 39% on grounding and 18% on numeric VQA with under 10s inference latency on an A100 GPU

• Developed a full-stack canteen automation platform using Django, React.js, TypeScript, and PostgreSQL
• Implemented secure role-based access control, OTP email authentication, and a digital wallet with cryptographic PIN verification
• Integrated asynchronous WebSockets through Django Channels for low-latency, real-time order updates on the manager dashboard
• Built live menu management with CRUD operations for owners and an algorithmic wait-time estimator driven by the order queue

• Implemented a fast sMRI preprocessing pipeline for 85+ GB of ADNI sMRI scans to improve high-dimensional feature reliability
• Designed and benchmarked standard CNNs, 3D CNNs, attention-based models, and dual-branched CNNs for Alzheimer's disease detection
• Achieved 85% accuracy, competitive with modern SOTA benchmarks
• Leveraged supervised transfer learning on binary classifiers to detect early-stage disease progression (MCI)

• Built a cost-effective 360-degree perception system that replaces LiDAR with monocular cameras
• Developed a Genetic Algorithm-based localization method from camera observations
• Fused it with odometry for real-time localization at 20 Hz
• Integrated YOLO11n + Geometric Ray Casting for 3D coordinates
• Achieved <10 cm positioning error without active depth sensors

• Trained YOLO11n with a Kinect depth camera for 3D ball tracking at 20 FPS
• Formulated trajectory prediction algorithm for interception point estimation
• Implemented custom velocity controller for goalkeeper movement
• Validated via custom 2D Python simulator before physical deployment

• Built a custom Tetris environment with Pygame rendering and video export utilities
• Implemented a Double DQN model and training pipeline for this environment