Himanshu
Dhurve

Developed citation-grounded RAG system for 580-page OpenFOAM documentation achieving 91% top-1 retrieval accuracy through metadata-aware cross-encoder re-ranking (+44% over baseline). Engineered end-to-end pipeline with multimodal query routing and hierarchical parsing.

Accelerated CFD simulation for Lid-Driven Cavity case by 95% by implementing Predictor-Corrector method on GPU. Engineered custom parallel processing solution using Numba CUDA kernels for high-performance computing.

Fine-tuned Qwen2.5-Math-1.5B-Instruct using Unsloth with compact MLP head to improve model interpretability of student misconceptions. Achieved 4.5% performance gain (0.885→0.925) on MAP leaderboard.

Developed video-based license plate detection system using fine-tuned YOLO11 and EasyOCR that eliminates text flickering through multi-frame stability buffer with majority voting algorithm across 12-frame rolling window.

Video instance segmentation using PyTorch's pre-trained Mask R-CNN model. Performs object detection and pixel-level segmentation on video streams, identifying and masking individual objects with colored overlays and bounding boxes.

Developed C++ solvers using Finite Volume Method (FVM) for 1D steady-state Convection-Diffusion equations with Upwind Differencing Scheme. Implemented Gaussian Elimination for tri-diagonal matrices with matplotlib-cpp visualization.

Led Aerodynamics Simulation Team for FSAE front wing design, performing 2D and 3D RANS CFD simulations in ANSYS Fluent. Used SolidWorks for design changes to optimize lift-to-drag ratio and downforce within SAE regulations.