Pipeline

Tools and techniques for moving performance capture from script to engine. Faster, cheaper, more consistent. Four years building proprietary pipeline at Deck Nine, followed by independent R&D extending the same thread into markerless capture and agentic authoring.


Deck Nine Pipeline

SceneCalculatorPro | Script Scoping and Auditing

Motion Capture Director · Tool Development · 2025

SceneCalculatorPro parsed .xml data from Deck Nine’s proprietary script editor and compared it against production scope, reference video shot on stage, and in-game cinematic minutes once implemented. Midway through Life is Strange: Reunion, SceneCalculatorPro identified a 35% project overage and became an essential tool in bringing the game back into scope and budget.

SceneCalculatorPro | Script Scoping and Auditing dashboard

Backfill | Automated Reference Video Assembly and Delivery

Motion Capture Director · Tool Development · 2025

Backfill is a tool suite that automated the assembly, editing, subtitling, and notation of reference videos from the mocap stage. It leverages the DaVinci Resolve and Confluence APIs to deliver witness video and performance context to downstream teams as a searchable, indexed Confluence library, turning a manual, day-long handoff into a push-button pipeline step.

Backfill | Automated Reference Video Assembly and Delivery

Narrative Engine

Scripty | branching narrative editor | state management engine

Scripty is a branching narrative editor built for interactive stories that accumulate hundreds of player choices without the exponential branch explosion of traditional tools. Each decision writes to a running state portrait of the player, letting scenes react to patterns of behavior instead of single flags. A CodeMirror screenplay surface sits alongside a live state inspector, giving writers and directors one authoring environment for drafting, auditing, and testing narrative state across the full runtime.

Scripty: branching narrative editor with state management

A dedicated wire-planning and whiteboard layer turns the same document into a visual workspace: scenes become drag-and-drop cards, choice dependencies render as explicit wires between them, and story beats can be rearranged spatially before a single line of dialogue is touched. The two surfaces are bi-directional. Build the wire plan first and let it scaffold the script, or start writing and generate the wire plan from what you’ve already drafted. Planning and writing stay in the same tool, anchored to the same underlying state.

Scripty: wire-planning and whiteboard layer

See full project →


Markerless Capture

bodypipe | video-to-mocap

bodypipe is a markerless pipeline that extracts body, face, and hand data from video: a machine-learning-driven replacement for the Vicon and Faceware systems of old. Running on consumer GPUs, with a Qt GUI, real-time 3D preview with Live Link connections, and BVH/FBX export.

bodypipe: markerless body, face, and hand capture from video
bodypipe: multi-person capture with identity tracking

See full project →


facepipe | tracking refinement and regression model training

facepipe takes raw MediaPipe face tracking and refines it into performance-ready ARKit blendshape data. Actor-specific profiles capture corrections through Cubase-style automation handles, letting a director apply, undo, and reshape tracking curves directly on the timeline. Those corrections feed regression models trained per-actor, so the system sharpens to the performer the more you work with them. Live Link output streams straight into Unreal.

facepipe: tracking refinement and regression model training

sandpipe-body | inertia and physics from monocular video

sandpipe is a general-purpose pixel-physics sandbox: elemental particles with unique physics impacting a spring system and boundary collisions, running on the GPU. sandpipe-body is the pipeline that extracts physics cues from monocular video. The research roadmap folds sandpipe-body back into bodypipe’s skeletal data as a physical prior. The result will be mocap with weight and grounding baked in, recovered from a single camera.


Agentic Network

The pipeline runs on an interconnected homelab of heterogeneous agents: an orchestration model directing a cluster of specialized local LLMs, all of them reading from and writing to a single shared memory layer. Tailscale holds the nodes together as a private mesh, and a central dispatcher routes work to whichever model is best suited, long-context reasoning, code, vision, embeddings, without ever leaving the local network.

A unified memory service gives every agent the same persistent context: semantic memories that consolidate over time, heat-decayed chat memory that cools when topics go quiet, and raw transcripts of every session ingested automatically. Inbound traffic flows through a Telegram bridge that carries conversations from phone to orchestrator. Outbound messages are council-gated before anything leaves the network. Autonomous wake cycles run on cron, checking state and picking up work while I’m away.

Agentic Network Topology Conceptual topology of the homelab agent network. A central orchestrator connects to Memory, Dispatch, Inbound, and Outbound layers. ingests gates reads / writes routes Inbound Telegram bridge wake cycles Memory semantic memories heat-decay chat session transcripts unified API · Tailscale Oracle Network Orchestrator Opus · 1M context Dispatch qwen3-next-80b gpt-oss-120b qwen3-coder-next vision · embed Outbound council-gated messaging scheduled triggers