I don't just direct the work. I build the systems that produce it. Custom AI pipelines, trained models, and node-graph workflows that turn craft into something repeatable, consistent, and fast.
Flagship system
Reelly
A fully automatic, self-improving video engine that turns raw single- or multi-cam footage, images, or even just text into fully finished long-form and short-form videos, ready for content platforms.
Reelly edits itself and improves itself. It understands the material, directs the cut, and assembles finished deliverables (rendered videos plus a DaVinci Resolve project with separate music, SFX, caption, and title tracks). It runs its own QC gates, produces platform-tuned short cuts alongside the long-form edit, and learns from how each video performs to reshape the next one. Built in Python, orchestrating Gemini, fal, ElevenLabs, and Whisper, with a firm human-publishes-always rule.
Problem. Series was scaling fast across internal work, partnerships, and Netflix titles. Trial-and-error prompting couldn't hold consistency or keep pace.
What I built. Modular ComfyUI node-graph pipelines: four production paths from exploration to final, a custom avatar system generating six consistent character emotions, and bot-assisted batch video generation. Emotions, consistency, and production logic live inside the workflows themselves.
Result. 100,000+ assets shipped at 180× speed across multiple Netflix titles and Lorespark. It became ComfyUI's first official customer case study.
Problem. Keeping characters and environments on-model across thousands of generations and branching narratives.
What I built. Trained and fine-tuned LoRA models for style-consistent characters and environments at production scale.
Result. The same character stayed recognizable everywhere. Consistency held across branching stories, not just single shots.
FaceMesh Placement Checkpoint
Problem. In a scalable avatar system generating thousands of objects, heads drifted. Position and internal rotation wouldn't hold.
What I built. A workflow checkpoint using FaceMesh to lock a reference point between the eyes, tracking rotation and placement, and measuring face shape and feature spacing along the way.
Result. Killed the drift, and the extra measurements compounded into higher overall consistency.
Patolli: Building Where No Tool Exists
Context. Solo-developing Toxtli & The Moon Jar (Steam) needed production-grade tooling no off-the-shelf product offered.
What I built. In Unity/C#: procedural level generation, ML-Agents reinforcement-learning QA agents, custom automated test runners, and MCP-driven editor automation.
Result. Proof of tool-builder, not tool-user. I build the solution when the tool doesn't exist.
