Chovadé
Independent R&D · Local AI

Special projects.

Not every Chovadé project ends as a film. Some are tools—local AI systems for creative work, production orchestration, and agent automation.

These projects explore how models can operate privately, visibly, and under human control.

Local

Models and media stay on owned hardware.

Private

No third-party service in the production path.

Inspectable

Work, decisions, and failures remain visible.

Human-approved

Authority stays explicit at meaningful steps.

Active developmentLocal-first

SAM

Local-first AI agent system

A computer-use and coding-agent environment that brings projects, tools, multimodal vision, research, memory, browser interaction, and scheduled work into one governed system.

  • ControlScoped authority, explicit approvals, budgets, and stop conditions.
  • EvidenceVerification, durable state, recovery, and inspectable action history.
  • RangeCoding, computer use, research, vision, speech, memory, and proactive workflows.

Qwen-Agent · MLX · Python · FastAPI · React · TypeScript · SQLite

SAM local agent interface in its ready state
SAM · Agent interface
Work in progressCreative systems

Blackfire

Agentic creative framework for film and television

A local-first production system that maps development, pre-production, generation, review, editorial, audio, colour, and final assembly into one inspectable workflow.

  • PipelineStory bibles, screenplays, shot lists, references, storyboards, and animatics.
  • ProductionVersioned assets, render queues, review stages, and final delivery.
  • GovernanceHuman approval gates keep creative and production decisions accountable.

LTX-Video · Qwen · MLX · FastAPI · SQLite · FFmpeg · Apple Silicon

Blackfire character and environment development board
Blackfire · Development board
Private current buildOpen-source release consideredLTX-2.3

LTX-2-MLX BX

Apple Silicon inference rebuild

An MLX-native rebuild of the LTX-2 inference path for Apple Silicon. Steve forked the original port, then spent weeks tracing the PyTorch reference, rebuilding execution boundaries, testing media parity, and optimizing the pipeline without hiding failures behind fallback paths.

  • Parity97%+ correlation on the measured validation path, backed by media-level checks.
  • Performance1.86× faster than the prior production MLX quality route on a matched benchmark.
  • MethodPrepared weights, MLX-native execution, granular profiling, and explicit acceptance gates.

LTX-2.3 · MLX · Apple Silicon · Python · FFmpeg · PyTorch parity

View the public fork ↗

The repository is an earlier snapshot; the current local rebuild is substantially ahead.

A full open-source release is being considered. LTX-2.5 support is a possible future update, not yet committed.

Cinematic night scene generated through the LTX-2-MLX BX pipeline
LTX-2-MLX BX · Cinematic output

More work in progress.

This page will also hold smaller creative tools, experiments, and systems as they become ready to show. The common thread is practical AI built around privacy, authorship, and clear human control.

Discuss a special project