AI Engineer
I build the agentic systems behind generative film: multi-agent pipelines on Azure AI Foundry that turn a script into finished frames, the eval harnesses that catch their mistakes, and the inference that runs the whole thing. It's in production right now on streaming series and upcoming features.
Real frames behind, from three productions.
India's first AI-generated premium series, now past 26.5M views. I built the multi-agent shot pipeline behind it and the eval harness that gates every stage, running 50 to 200 generations a day at peak with 50+ artists downstream.
Hansal Mehta's AI cooking show. I wrote the resumable runner that carries an episode's script to finished shots, 88 generated Bernini prompts for Ep04 alone, so a failed render retries without losing the batch. I also built the audio side: ElevenLabs generation and voice cloning, so one cloned voice carries the show across several languages instead of re-recording per language.
Storyboards and the image-to-3D pipeline for an AI-generated theatrical feature. Trellis meshes land at 1M+ faces with no materials, so they run through headless Blender retopology and texture baking before anything reaches a scene.
Storyboard generation and style transfer for an upcoming feature. Character LoRAs trained on Flux and SDXL with reference-image control hold likeness across hundreds of shots, which prompt text alone will not do over a film-length run.
A photoreal interactive avatar built on the studio pipeline: motion capture, isolated reference voice, and multi-character lip-sync held in sync across a full clip.
Six hosted agents carry a script to finished shot panels. Each hands a typed artifact to the next and is versioned in a registry, so a bad hand-off fails loudly instead of quietly corrupting the run. Tuning the judge was the real work: at threshold 0.5 it over-rejected pencil sketches for style bias and hallucinated absences. Dropping to 0.40 and telling it plainly that style is fixed cut the false rejections.
Script inShot panels out
Each agent is versioned in a registry, so a bad hand-off fails loudly instead of quietly corrupting the run.
A browser tool for VFX teams: segment objects, matte hair and fur, remove anything from a shot. I A/B-tested SAM 3.1's tiled video mode against SAM 3, and it is 20x faster but only reaches 54% recall on dense crowds, so it ships for tracking up to about 100 subjects and SAM 3 stays the default for crowds. The models will not all fit in memory at once, so the backend lazy-loads and swaps one at a time.
Auto multi-ID
Stock ViTMatte weights inherit a research-only licence, which blocks external client delivery. I retrained an Apache-2.0 replacement that had to beat the model it replaced, and on held-out frames it does: SAD 9.86 against 14.61, with the flyaway hair intact. The trap worth knowing is normalisation, where ImageNet against (0.5,0.5,0.5) is roughly an 8x difference in error and hides as a training bug.
Two photos in, an explorable 3D world out: collider mesh, panorama and Gaussian splats, then any captured viewpoint re-rendered as a high-res still. Four model vendors sit behind one job-state API so a slow or failed stage retries on its own rather than sinking the whole job.
Controls a fleet of GoPros at once: BLE discovery, COHN live preview, synced record, and a parallel pull of every SD card. There is no hardware genlock on these bodies, so sync is near-simultaneous WebSocket commands corrected afterwards by embedded timecode. One codebase, installers for macOS, Windows and Linux.
Camera fleet
Turns a screenplay into a shot-by-shot storyboard and holds every character on-model from reference sheets. It runs on hosted models with no dedicated GPU, which is what makes it cheap enough to leave running for a whole film, and the artist-facing details that mattered were export, job pause and resume, and honest progress.
"Beyond the treeline, the forest opens onto a moonlit lake fed by a crashing waterfall. Through the silver mist, a woman stands drenched beneath the falls, singing, and Bhim, mace in hand, draws back into the shadows."
SHOT 30XWS
SHOT 50CU
SHOT 60MS
SHOT 90CU
SHOT 130CU
SHOT 150MS
SHOT 170CU
SHOT 300MSReal panels from one scene Cue storyboarded end to end: all 33 shots from the screenplay. Working from just the two reference sheets above, it holds Bhim and Hidimba on-model across every panel, whatever the framing.
An agent that writes Blender Python from a photo, renders it, and scores the render against the source with a vision LLM, feeding the diff back in until the reconstruction holds up. Generate, render, verify, refine.
Turns flat product photography into the frames a brand would publish: the real garment, on a consistent model, with colour, print and construction held to the reference. A customer receives the actual item, so a drifted print or a shifted shade is a return, and every decision below follows from that. Analysis and art direction run as separate calls at different temperatures, fine print is cropped and upscaled into its own reference, and nothing ships until it clears a 7-point QC pass.
Eight garments from one run. The inset is the product photograph that went in; the frame behind it is what came out. The model stays the same person across the whole catalogue.
Colour is the criterion that actually fails. Everything else clears 96% or better, which is exactly why the pipeline reads colour deterministically instead of trusting one creative pass.
Point it at a YouTube URL, a podcast, or a local file, and it isolates one speaker's voice: diarization, transcription, speaker-matching and enhancement chained into a single command. Feeds reference voices into downstream lipsync and dubbing work.
Diarization One real 5-minute run · 2 speakers · 108 segments
0:002:305:00
The dominant speaker is the one the pipeline keeps. The 29% that is neither is silence, music or overlap.
One command: a URL in, one clean speaker's voice out.
Product photo in, a finished vertical ad out: a synthetic creator holds the product, talks about it, and the caption burns in, with no camera, no actor and no shoot day. Firecrawl pulls real competitor ads for the category before a script is written, so positioning is grounded rather than invented, and ElevenLabs generation and voice-matching pick a voice for the persona instead of defaulting to one stock narrator. Job state runs through a Cloudflare Durable Object per video, so a render can resume instead of restarting a multi-minute pipeline from zero.
Available for hard, applied AI roles in generative media and production infrastructure. Fastest reply by email.
