Domain Expert × AI-Assisted Development
I build tools the way a production artist would - starting from the friction I feel in the actual work, not from a feature list. Years in the pipeline tell me exactly what to build and whether the result is genuinely right; AI-assisted development is what lets me ship it.
Turned hours of manual, error-prone asset auditing into instant, visual, confident optimization - adopted by the entire VFX discipline.
VFX artists spend a meaningful slice of every production cycle on work that has nothing to do with the actual craft: figuring out what an effect depends on and how expensive it is. Before the tool, that work was entirely manual.
Hunting through deep folder trees to find the textures, meshes, and sub-systems an effect referenced.
Opening files one at a time to trace what pointed to what, with no reliable way to know if an asset was still in use.
No direct line of sight into which assets were driving an effect over budget - optimization was reactive and imprecise.
Making a change, rebuilding, and re-checking by eye, then repeating until the numbers looked acceptable.
The cost to production was real. Optimization passes - a routine, recurring part of shipping effects under a memory budget - became a tax on the whole discipline. Effects shipped heavier than they needed to, and cleanup of unused assets was inconsistent.
Identifying the opportunity. The insight came from production experience, not from a feature request. Doing the optimization work myself, repeatedly, made the pattern obvious: the bottleneck wasn't the decision to optimize - artists knew how to make effects cheaper - it was the information gathering required before any decision could be made.
Designing around the artist workflow. I mapped the real day-to-day path an artist takes through an effect: open the system, find its emitters, trace each emitter to its meshes and textures, check whether those assets were shared or unique, and estimate the weight. The tool was designed to mirror that mental model rather than impose a new one.
Why AI-assisted development was the right call. The problem was well understood by exactly one kind of person - someone who lives in the production workflow. AI-assisted development closed that gap: it let the person with the domain expertise also be the person who ships, without waiting in an engineering queue or losing fidelity through a spec-handoff.
A single desktop application that sits alongside the artist's normal work and answers three questions instantly: what is this asset, what does it cost, and what depends on it?
Browse and search all asset types with visual thumbnails. Finding the right asset is immediate instead of a folder hunt.
Inspect meshes directly in the tool with geometry stats visible - no round-tripping through a heavier application.
See exactly which systems use a given asset, and which assets a system pulls in. Makes it safe to identify and remove genuinely unused assets.
Surface the stats that matter for a memory pass so heavy contributors are obvious. Target the actual cost drivers instead of guessing.
Project-level reporting shows what's used, what's not, and by how much - supporting cleanup at scale, not just one effect at a time.
The workflow shifts from investigate-then-guess-then-verify to see-then-act.
I didn't hand the AI technical specs. I described problems the way an artist experiences them: "I need to see everything this effect depends on." The AI translated intent into implementation.
Conversational and incremental - describe a behavior, get a working version, use it, react, refine. The tool grew through many small, evaluated iterations.
I could look at output and immediately know whether a dependency count was right, whether a stat was meaningful, whether a workflow matched how artists actually work.
I made the higher-level calls - standalone desktop tool, single window, mirror the artist's workflow. The AI filled in the how; the shape was directed by product intent.
AI didn't replace engineering expertise - it made engineering expertise non-blocking for a problem that only a domain expert could correctly specify and validate.
The tool had very easy team adoption that helped us meet very demanding performance optimization goals under strict deadlines.
The information-gathering that used to dominate an optimization pass was largely eliminated, cutting the manual portion roughly in half.
Artists can now identify and fix memory-heavy effects confidently - optimization shifted from a reactive guessing game to a targeted, measurable action.
Time previously lost to manual auditing and bookkeeping went back into actual creative work.
As an early AI adopter, I helped inspire the Engineering team to dive into AI tools after demonstrating this tool, and influenced their own product goals with some of the features I developed.