The lab behind the work.
Open source is where I pressure-test multi-backend runtimes, GPU coordination, identity pipelines, plugins, and recovery paths without waiting for a roadmap. Active work sits beside stable systems and older references so the lifecycle stays explicit.
01 · AI & ML
Diffusion, LLM, and ML tooling. Where most of my recent open-source energy lives.
WebbDuck
Local-first multi-backend image studio
A self-hosted image studio built on FastAPI, Diffusers, and isolated runtimes. Supports SDXL, FLUX, Krea 2, and Qwen backends, capability-driven controls, cross-architecture Identity / Persona workflows, editing and upscaling, LoRA discovery, a searchable local gallery, and plugin integrations like DuckMotion and DNADuck.
DuckMotion
Wan 2.2 image-to-video plugin for WebbDuck
A local image-to-video plugin that adds staged uploads, model discovery, runtime checks, queued generation, cancellation, and a plugin-local gallery to WebbDuck. Built around Wan 2.2 and real local-GPU constraints.
DNADuck
Identity clustering and LoRA dataset prep
A local face identity extraction and clustering pipeline for image libraries. Scans folders, embeds faces with InsightFace, groups identities with persistent SQLite state, supports review/merge/search workflows, and exports LoRA-ready datasets with training hooks.
llm-stack
Retired local LLM infrastructure stack
A superseded shell-driven local LLM + retrieval stack retained as an implementation reference. It reflects an earlier generation of the lab and is no longer under active development.
02 · Tools & Automation
Small, focused utilities — the kind of thing that quietly saves a team a lot of time.