AI & Computational Systems

Designing AI-assisted pipelines across pre-production, real-time production, and post-production.

Systems-First Approach to AI in Production

I approach AI as a systems problem, not a standalone tool. My focus is on where AI meaningfully accelerates or augments real production workflows, from previs and environment generation to iteration speed, consistency, and handoff between departments. Every experiment I build is evaluated against production constraints: latency, controllability, repeatability, and integration with existing CG and VP pipelines.

OM: AI-Assisted Virtual Production Pipeline

This pipeline explored how AI can reduce iteration time in early look development while still feeding clean, controllable assets into a real-time VP workflow.

Director • AI-researcher

July 2025 - Present

Illustrated by Kira Narog

Production Designs by Sebastian Kapur

Directing and actively researching an AI-assisted virtual production pipeline aimed at improving previs and shot planning efficiency using limited real-world reference data. The system integrates Unreal Engine, LiDAR/photogrammetry, and AI-assisted Gaussian splat reconstruction (via tools such as World Labs) to convert small sets of stylized location imagery into metrically grounded, navigable previs environments. This workflow supports early-stage camera blocking, environment validation, and creative alignment across departments when traditional scouting or full photogrammetry is impractical.

This research examines how AI-driven spatial reconstruction can complement traditional previsualization workflows without replacing established CG or VFX pipelines.

AI Style Transfer & Consistency Research

ComfyUI · Custom LoRA Training · Python-Assisted Workflows

July 2025 - September 2025

This research focused on understanding how AI-driven style transfer behaves across repeated iterations and asset contexts, rather than producing final imagery. Using custom-trained SDXL LoRAs and ComfyUI-based workflows, I explored the limits of stylistic consistency when AI outputs are introduced into a production pipeline. The goal was to identify where AI can accelerate early look development without compromising downstream continuity, scale, or asset reliability. These findings informed later decisions about where AI should remain an assistive tool rather than a generative authority within Virtual Production workflows.

Rather than treating AI outputs as final assets, this work reframed them as directional inputs; valuable for ideation, reference, and early-stage validation, but requiring careful translation into controllable CG or real-time environments. This approach helped define guardrails for AI usage that preserve artistic intent while maintaining technical predictability in collaborative pipelines.

Independent research into AI style transfer and consistency using custom-trained LoRAs and ComfyUI pipelines to understand limitations of AI outputs in production contexts.

This research directly informed how AI outputs can be safely integrated into pipelines without breaking continuity or downstream workflows.

AI-Assisted Interactive Homepage System

Real-Time Scroll-Driven Visual System · AI-Integrated VAD Workflow

Creative Technologist · Virtual Art Department (AI Integration)

August 2023 - December 2024

This project explores how AI-generated imagery can be responsibly integrated into real-time, interactive web experiences without compromising spatial accuracy, performance, or authored intent. Using custom drone footage as the physical environment reference, I integrated a peer-designed, production-ready train asset via Higgsfield AI, ensuring scale, lighting, and motion continuity aligned with the real world. The resulting sequence was translated into a scroll-driven image system that transitions seamlessly into original film footage, reinforcing continuity between AI-assisted content and traditional production pipelines.

Original IP train design by Jesse Van Norman.

Designed a scroll-driven image-sequence system for real-time visual storytelling on the web, optimized for performance, preloading, and perceptual smoothness.

Higgsfield AI · Virtual Art Department Workflows · Custom IP Integration · Scroll-Driven Interaction Design · Web-Based Real-Time Playback