Hruthik Mula

Researched AR/VR and hand-gesture systems to improve industrial assembly/disassembly performance for Meta Quest 3. Built a model-agnostic Unity automation pipeline to convert FBX + JSON instructions into complete VR assembly scenes, including auto-generating targets, grabbable parts, colliders, rigidbodies, labels, and progress UI. Implemented a digital-twin registration pipeline in Python using Open3D point-cloud sampling (from CAD/STP), feature descriptors (FPFH), coarse alignment (RANSAC), and ICP refinement to align virtual overlays to real hardware. Integrated MediaPipe hand tracking (21 landmarks) with path smoothing and PCA-based direction estimation to infer insertion directions, and contributed to an LLM instruction engine (fine-tuned GLM-4-9B with 4-bit QLoRA) producing validated JSON guidance text.

Delivered an end-to-end adaptive VR skill assessment system within a 75-day internship. Built from scratch three interactive VR stages assessing reflexes, aiming, reaction time, and decision-making, with real-time adaptive difficulty driven by live scoring data. Created custom 3D models and optimized environment assets in Blender within Meta Quest 2/3 performance budgets, using AI-assisted tooling (Claude, Codex, Meshy 3D) to accelerate asset generation and resolve complex XR interaction issues. Implemented VR interaction layers (locomotion, object interaction, spatial UI) and performance tracking metrics for accuracy, shooting precision, and movement behavior classification.

Researched and built an AR/VR/MR pipeline for hand-gesture-assisted industrial assembly and disassembly performance, targeting Meta Quest 3. Developed a model-agnostic Unity automation system that converts FBX models plus JSON instructions into full VR assembly scenes, including snap targets, grabbable parts, colliders, rigidbodies, labels, and progression UI. Implemented digital-twin registration in Python using Open3D point-cloud sampling from CAD/STEP parts and performed coarse alignment via FPFH + RANSAC with ICP refinement to align real hardware overlays to virtual models. Integrated media pipeline hand tracking with landmark smoothing and PCA-based direction estimation, and contributed to an LLM instruction engine (GLM-4-9B fine-tuned with 4-bit QLoRA) producing validated JSON guidance text.