About
I work at the seam between two crafts that rarely speak to each other. By day I am a compositor at Industrial Light & Magic in Sydney, integrating visual effects into live-action footage at the precision a feature film requires. Outside that work I build generative AI tools that try to do what compositors actually need from them — produce scene-referred linear EXRs, respect color management, behave like the rest of the pipeline behaves. The two halves of the work look unrelated from the outside. From the inside they are the same craft asked at different scales.
The path here started in Bangalore in 2016, doing junior compositing at MPC. The years that followed cycled through Wonder Woman, Justice League, The Greatest Showman, Detective Pikachu — the kind of franchise calendar where you learn the craft across a wide range of shot types because the schedule does not let you specialize. In 2021 I moved to Luma Pictures in Melbourne for Marvel work. In 2022 I joined ILM in Sydney, where I am now, working on Mission: Impossible — Dead Reckoning, Indiana Jones and the Dial of Destiny, The Rings of Power, Wicked. Thirty-plus feature credits to date, with more in flight.
The pivot toward AI was not a deliberate career move. It started as a question I could not answer with the tools I had: why does diffusion-generated content fall apart when you put it in a comp? The honest answer turned out to be that the model does not know what scene-referred light is — that the gap between "generates pretty images" and "generates pixels a colorist can grade" is wider than most of the AI tooling community appears to realize, because the people building the tooling have rarely sat in front of a Nuke graph at midnight trying to fix the highlight rolloff on a hero shot. Closing that gap, in stages, became the through-line of the work documented on this site.
The pattern I keep returning to is that the meaningful design decisions live in the seam itself — in the small choices that only a person who cares about both sides would think to make. The architectural decision to extend a VAE's channel count to carry both tonemapped and linear representations came out of a compositor's instinct about pixel alignment, not an ML researcher's. The decision to hardcode an OCIO config rather than load it dynamically came from too many years of debugging color-management-on-Friday-afternoon problems. The choice to invent a loss term rather than reach for more data came from understanding that what colorists call acceptable is not what benchmarks call acceptable. None of these decisions is novel in isolation. The combination — caring about both crafts at the level of detail each requires — is uncommon enough to be useful.
Most of what is documented on this site is openly licensed and publicly available. A small amount cannot be released for contractual rather than strategic reasons. My default with research-adjacent work is to publish; the parts that are private are private because they have to be, not because I think the work is too valuable to share. The HDR generation pipeline, the Nuke nodes for ComfyUI, and the ComfyUI MCP server are all open source and in use by other people, which is the outcome I prefer.
I am open to research and engineering roles in generative AI for film, VFX, and adjacent media. The work that interests me most is at companies building models that need to behave inside professional production pipelines — Black Forest Labs, Runway, Luma AI, Adobe Research, NVIDIA, Foundry, the AI groups inside Wētā FX or Wonder Dynamics, and the smaller labs and startups circling the same problem set. The right role is one where the model is expected to ship into a real workflow rather than a benchmark, and where the team values the bridge between traditional production craft and modern ML infrastructure as something more than a job title.
If any of that resonates, the contact section on the home page has the relevant ways to reach me.