ALEX MOED Computer Vision Engineer U.S. Citizen | Canadian Citizen | Boston, MA | Willing to relocate alexmoed@gmail.com | (617) 934-3774 | alexmoed.com | linkedin.com/in/amoed SUMMARY ------- Computer Vision Engineer with 13 years in VFX production as a 3D artist, plus a Master of Science in AI for Media (Highest Honors). Specialization: 3D reconstruction, Gaussian splatting, object identification and segmentation, and diffusion models. Built end-to-end systems that turn room-scale Gaussian splat scans into searchable 3D asset libraries, with each object segmented and described in open-vocabulary format. VFX credits across DNEG, Digital Domain, and MPC, including Dune: Part Two (Academy Award) and The Last of Us (Emmy). TECHNICAL SKILLS ---------------- Languages & ML: Python, PyTorch, CNNs, generative AI, vision-language models (VLMs), diffusion models, Hugging Face Transformers, fine-tuning (LoRA) CV / 3D / Tools: Gaussian Splatting (3DGS), 3D reconstruction, COLMAP (SfM), Open3D, OpenUSD, instance segmentation, semantic segmentation, object detection, open-vocabulary detection, 2D segmentation models (SAM3), local LLM inference (vLLM), GPU-accelerated inference, model quantization, Docker, AWS, ComfyUI, Houdini EXPERIENCE ---------- Computer Vision Engineer - Volustor, Remote (UKRI research grant) Sep 2025 - Mar 2026 - Designed an agentic pipeline that turned multi-camera volumetric captures into searchable libraries of identified, tagged 3D assets. This addressed the manual-cleanup bottleneck that has limited production use of volumetric scans. - Built a system that segmented objects in room-scale scans and generated open-vocabulary descriptions for each asset. - Built an embedding-based system that matches the same object across non-sequential multi-camera views, cutting duplicate assets without requiring temporal continuity between frames. - Extracted individual objects as standalone Gaussian splats from room-scale scans using visual hull carving with 2D segmentation masks. - Reduced AWS compute costs by quantizing the larger models and tuning quantization levels to preserve segmentation and detection quality. - Deployed the full pipeline as a containerized workload on AWS, with PostgreSQL and pgvector backing the searchable asset library, taking a research prototype to production infrastructure. - Helped build Volustor's spatial fingerprinting system that detects unsanctioned reuse of performer identities by matching 2D images of people against studio-captured Gaussian splat scans. Recognized as a top-4 finalist in Innovate UK's inaugural Agentic AI Pioneers Prize. Senior Environment Generalist TD - DNEG, Vancouver, BC Feb 2022 - Feb 2024 - Dune: Part Two (Academy Award winner): Optimized environment assets for the Arrakeen Basin. Designed a procedural system for background dune generation, including the tracks and impressions that vehicles and characters left in sand. - The Last of Us (Emmy winner, Best VFX in a Series): Drove procedural building destruction systems to generate collapsed structures, hand-adjusted sections to art direct the final look, and built a custom system for generating destroyed beams that could be scattered across shots. - Fighter: Reconstructed real-world mountain landscapes as photo-real VFX environments using Google Earth terrain data as the base. Rebuilt the geometry to production quality and combined hand-remodeled featured sections with procedural systems in Houdini. Added high-detail geological features that integrated seamlessly with the wider landscape. - Shazam! Fury of the Gods: Generated procedural vines that engulfed buildings using hand-drawn guide curves. Senior Environment Artist - Digital Domain, Vancouver, BC Oct 2020 - Feb 2022 - Doctor Strange in the Multiverse of Madness: Modeled featured buildings, authored procedural scatter and fracture systems, ran ground look development, and mentored junior artists. - Loki: Built bridges and city structures, laid out scenes, and generated procedural rock and terrain. - Black Adam: Ported legacy environments into an OpenUSD-based Solaris pipeline, adapting models and textures for a new workflow. Lighting Artist - Goldtooth Creative, Vancouver, BC Jul 2020 - Oct 2020 - Marvel Realm of Champions: Lit and composited characters and environments for a video game trailer. Environment Artist - Digital Domain, Vancouver, BC Oct 2019 - May 2020 - Free Guy: Modeled, textured, and developed looks for interior and exterior environments. Environment Artist / Lighting Artist - MPC, Vancouver, BC Feb 2018 - Oct 2019 - Sonic the Hedgehog: Lit and rendered large-scale environments, characters, and vehicles. Textured a full CG highway. - Detective Pikachu: Laid out featured sections of Ryme City, organized assets, and handled modeling and look development. - Maleficent: Mistress of Evil: Designed lighting for the armory sequence across full CG environments and set extensions. - Skyscraper: Ran look development and render optimization for 50-plus animated frosted glass panels. 3D Artist & Freelance Generalist - Various studios (Chicago, Los Angeles, Vancouver) Aug 2011 - Sep 2017 - Studios: Odd Machine, The Mill, Hoax Films, Bardel Entertainment, Comen VFX, and Entity FX. Shipped commercial, broadcast, and episodic VFX across multiple pipelines. - Designed and built an in-house 45-server render farm at Odd Machine and upgraded supporting file server infrastructure. - The Vampire Diaries: Mesh-tracked live-action actors using proprietary tools and match-moved 3D models into live plates for episodic television. - Led client meetings, bid preparation, rigid body simulations, and CAD-to-Maya workflows for industrial design clients. EDUCATION --------- MSc Artificial Intelligence for Media - Bournemouth University, UK 2024 - 2025 - Grade: Distinction (Highest Honors) - Program: A research-focused conversion course at the UK's National Centre for Computer Animation, designed to move media professionals into machine learning. Coursework in Machine Learning for Media Production, Data Mining on Multimedia Data, Software Engineering for Media, and Media Data Analytics and Modeling. - Group Project: Built a procedurally-generated city for autonomous vehicle simulation training. - Masterclass (Industry partnership with Volustor): Surveyed current Gaussian splat segmentation methods. Built SAM2-based pipelines for 2D video segmentation, and created a pipeline that used language-model-predicted 3D bounding boxes to classify points and export labeled point clouds via Open3D. - Thesis (SegSplat): Goal: Test whether an RGB-D-trained 3D instance-segmentation model could generalize to Gaussian splats, for which no labeled datasets exist. Method: Paired Meta's frozen, pretrained Sonata encoder with a trainable PointGroup head, trained on ScanNet v2 with targeted augmentation to bridge the domain gap. Result: 64.4% mAP@0.5 on ScanNet v2, above the 63.6% baseline; generalization to splats confirmed visually, with no labeled splat dataset for quantitative scoring. BFA Visual Effects - Savannah College of Art and Design (SCAD), Savannah, GA 2007 - 2011 - Trained as a 3D generalist in lighting, texturing, and compositing - SCAD Portfolio Scholarship recipient SELECTED PROJECTS ----------------- ELITE Diffusion Image Enhancer - Portfolio project (2026) - Goal: Restore fine facial detail lost in degraded Gaussian splat avatar renders. - Method: Fine-tuned SD-Turbo for single-step image-to-image restoration with LoRA adapters and skip connections, freezing the base so only 18M of 860M parameters train, on a combined L1, LPIPS, and style loss. - Results: LPIPS 0.205 across 600 held-out images at 80ms per image; deployed as a live Hugging Face Spaces demo. Agentic Gaussian Splat Object Segmentation - Self-directed research prototype (2026) - Goal: Autonomously decompose a Gaussian splat scan of a room into a clean standalone splat for every piece of furniture and decor, plus a separated background. - Method, Step 1: A local vision-language model (Qwen 3.6, served with vLLM) inventories the scene from rendered top-down and eye-level views, selects detection categories by room type, and routes every object it finds into extraction. - Step 2: Each object is isolated in 3D by back-projecting its SAM3 masks from many camera views, keeping only the splats confirmed by a consensus of those views, then removing the floor and any leftover points that fall outside the object's shape. - Step 3: An automated QC gate reviews each object's 360-degree renders and writes an open-vocabulary name and description into a manifest. Packaged as a single Docker image with FastAPI, Open3D, and gsplat, plus an interactive Three.js viewer. - Advantages: No COLMAP, camera poses, hand-labeling, or per-scene prompting required. Synthetic 3D Instance Segmentation Dataset - Self-directed (2026) - Goal: Build an end-to-end generation pipeline to produce training data for 3D instance segmentation. - Method: LLMs write room descriptions, FLUX turns them into product images, and TRELLIS 2 converts images into meshes. Results to date: Over 6,200 GLB assets generated. - Advantages: Addresses the shortage of labeled 3D data that limits segmentation research. ---------------------------------------------------------------------- alexmoed@gmail.com | (617) 934-3774 | alexmoed.com | linkedin.com/