CS undergrad at PES University building LLM-driven systems and publishing computer vision research. I turn messy, real-world data into models that actually perform — across NLP, deep learning, and healthcare AI.
Multimodal framework fusing RF features (AoA, TDoA, RSSI) with an LLM reasoning layer to classify 7+ UAV swarm formation types. Fine-tuned Mistral-7B-Instruct with 4-bit QLoRA (NF4, r=16 / α=32) on dual-T4 GPUs to generate natural-language analyst summaries from RF telemetry.
View on GitHub →Deep learning pipeline classifying haze density (light/medium/heavy) and enhancing degraded images. Achieved ~22% haze reduction and ~15% image quality improvement. Published at AIDE 2026.
View on GitHub →Time-Aware Transformer (EfficientNet-B4 backbone + temporal attention) modeling irregular clinical visit intervals on 21k+ longitudinal retinal images. QWK 0.9494, with reproducible evaluation across DDR, APTOS & MESSIDOR. Targeting OMIA @ MICCAI / ISBI 2026.
View on GitHub →Infrastructure-based multi-robot navigation: an overhead camera and edge compute localize ground robots to centimeter accuracy via ArUco detection, camera calibration & homography, then stream BLE drive commands. Built on ROS2 Humble with a clean two-node perception/planning contract.
View on GitHub →Normalized (3NF) relational schema with role-based access control and optimized SQL queries for crime-record retrieval, data integrity, and case metadata management.
View on GitHub →Presented a preprocessing methodology for non-homogeneous real-world hazy images from the RTTS benchmark dataset that significantly improves DehazeFormer's visual output — demonstrating ~22% haze reduction and ~15% overall image quality gain in autonomous driving scenarios.
Open to AI/ML internships, research collaborations, and anything involving large language models doing actually useful things.