Nov 2024 – Present
Technical Director — Applied AI & Product Engineering
I joined Orbital Vision in an Applied AI capacity, initially focused on image-generation model training and prototyping, before moving into technical leadership of OV25 — Orbital Vision's AI, 3D configurator, commerce, and developer-platform product. Across the role the throughline has been the same: take processes that were historically manual and expert-only — 3D configurator setup, material authoring, QA, asset discovery — and turn them into safe, agentic, AI-assisted systems that the wider team can actually use.
Go BackI post-trained an internal asset-generation tool on Stable Diffusion 3.5, developing advanced image-generation workflows with depth and Canny ControlNets, prompt conditioning, reference-image control, denoise tuning, and guardrailed review loops focused on product imagery quality. The goal was reliable, brand-accurate product imagery — not just pretty pictures — so the work was as much about control and constraints as it was about the base model.
I extended this into AI-based material and texture generation, trained on Orbital Vision's 10-year archive of V-Ray materials. That meant translating high-fidelity, offline-rendering material knowledge into WebGL-compatible PBR textures, shader constraints, and real-time browser configurator materials — bridging the gap between studio-grade rendering and what runs live in a customer's browser.
To make generation quality measurable rather than subjective, I built automated evaluation frameworks for image generation. These ran thousands of generated workflow variants across ControlNet strength, denoise steps, sampling algorithms, prompt structures, and reference-image strategies, then used model-based assessment to score visual fidelity, product accuracy, brand alignment, artifacts, and commercial usability. This turned “does this look right?” into a repeatable, comparable signal that could guide both the models and the workflows around them.
A large part of the role was leading AI-assisted creation of web-based 3D configurators — automating a historically manual process complicated by inconsistent 3D model structure, mesh naming, material setup, product logic, pricing rules, and customer-specific requirements.
I developed agentic configurator-setup patterns where users reference @materials, @products, @pricing, and @configurator, describe their requirements in natural language, and generate material mappings, visibility logic, option dependencies, and pricing conditions — with review-before-apply changes so nothing lands unchecked. This replaced complex UI that previously required heavy staff training and enormous amounts of time with plain human-language commands like “please update this organisation in this way,” backed by sandboxed, safe agentic skills and tools.
Underpinning that automation, I built a multimodal 3D asset-understanding pipeline that renders geometry, vertex data, and mesh-level components from multiple viewpoints, then uses Meta vision models and Claude Opus to infer component purpose, material groupings, and product semantics. In other words, it lets the system look at a raw 3D model the way a human technical artist would — and reason about what each part is for — which is what makes AI-assisted configurator generation possible despite messy, inconsistent source files.
I built a custom high-resolution WebGL/MSAA capture pipeline for OV25 configurators, rendering live React Three Fiber scenes into offscreen WebGL2 render targets at 4K/8K with GPU-aware multisampling. I then used these render pipelines to automate our 3D QA process — saving hundreds of hours — by assessing incorrect UV mapping, floating, colliding or misaligned meshes, and other visual imperfections.
Because this had to run reliably under heavy workloads, I designed the capture system for operational safety: persistent browser pools, memory cleanup through the Chrome DevTools Protocol, worker queues, retry budgets, screenshot fallbacks, graceful shutdown, SQS visibility extension, and recoverable task state in DynamoDB. It behaves less like a script and more like a resilient distributed service.
A decade of company knowledge was scattered across multiple NAS drives, cloud storage, separate email accounts, and internal notes. I centralised it into a single indexed, embedded, searchable library — making 3D roomsets, models, and other assets discoverable by reusing the render pipeline to convert 3D files into rendered images, embedding them, and adding them to a unified company “brain.”
On top of that, I designed an agentic workflow that goes from a client brief, to a lookup in the company brain, to finding a similar-brief roomset or 3D model. Where the brief needs reference accessories — a lamp, say, or other non-hero background items that can be slightly lower quality — it generates them on the fly using an image → multi-view synthesis → LRM → surface extraction → mesh/texturing pipeline inspired by NVIDIA research, then sends the close-match roomset, prebuilt accessories, and the original brief to our 3D artists to halve their work.
Across all of this, I designed and shipped production multimodal AI workflows spanning Claude/Anthropic APIs, Gemini, Veo, OpenAI, Meta vision models, Flux, and ComfyUI — including async job orchestration, prompt expansion, image troubleshooting, retry/failover handling, token ledgers, refunds, safety checks, and usage analytics. These are the operational concerns that separate a demo from a product that customers depend on.
I led a team of developers across OV25 implementation, combining architecture direction, pair programming, implementation support, and technical review to help engineers ship complex agent, configurator, pricing, and ecommerce features with confidence. I established reusable patterns, clarified ambiguous customer requirements, unblocked difficult integrations, and helped the team turn repeated implementation problems into stronger platform abstractions.
Go Back© 2026 Ziggy Baker