SivaNagendraSavarapu

A conversational AI bot that lets you interact with USD Stage using natural language. Ask questions about your scene graph, query prims, modify properties, and traverse hierarchies without writing a single line of USD Python. Built on OpenUSD + LLM tool-calling. 31 GitHub stars.

An end-to-end agentic pipeline built in C++ inside Unreal Engine that constructs 3D scenes from a single user prompt or image. Combines Claude (Anthropic Agent SDK) with UE MCP, Gemini/Meshy/Tripo for mesh generation, Depth Pro for depth estimation, Grounding DINO for object detection, and SAM2 for segmentation. Mesh placement via custom raycasting + GLTF plugin. Designed to shorten the loop between concept intent, scene assembly, and in-engine iteration for production teams exploring AI-assisted world building.

Designed and deployed Model Context Protocol servers for Maya and Unreal Engine, integrating Claude Code into Epic's live Fortnite asset pipeline. Among the earliest production MCP deployments at a major game studio, enabling LLM-driven pipeline automation across Modelling, Rigging, and Animation departments. The work focused on making AI assistance reliable inside real studio workflows, with tooling that respected production constraints instead of living only as a prototype.

Contributed to GenStudio, an internal generative AI platform integrating Gemini, OpenAI image models, Meshy, Tripo, and Rodin mesh generation for AI-assisted 3D asset creation. Also implemented C2PA content provenance workflows ensuring cryptographic credential tracking for all AI-generated image outputs. This work helped connect model experimentation with safer, traceable, and production-aware tooling for internal creative teams.

A RAG-powered assistant that auto-generates visual workflow graphs (analogous to Unreal Engine Blueprints) from natural language queries. Custom RAG backend built on internal pipeline documentation, enabling artists and TDs to describe a workflow in plain English and receive a structured, executable graph. The goal was to turn buried knowledge into something operational, reducing the gap between documentation, onboarding, and everyday pipeline execution.

Applied Latent Consistency Model research directly inside Houdini, enabling rapid AI-driven image generation within a production DCC environment. Explored how LCMs could accelerate iterative look development and concept rendering workflows without leaving the 3D application.