project-diffusion-workbenchin development
Projectmodel toolingLocal Diffusion Workbench
A working environment for dataset preparation, diffusion-model training, checkpoint review, and generated-output inspection.
Personal AI / Software Laboratory
Meadbee Labs
Building AI systems from the model up. Autonomous agents, models trained from scratch, generative tools, developer systems, and experiments with local AI.
Local firstContent drivenStatic by defaultOpen notebook
01 / Flagship system
Meadbee AI Harness
A local-first autonomous development environment centered on director and worker agents, extensible plugins, tool use, and long-running Wiggum loops.
The harness supports software work and model-driven creative workflows including Blender, image generation, and visual-novel planning.
ORCH-01Director
EXECWorkers
EXTPlugins
IOTools
ITERATEWiggum loops
02 / Featured projects
Systems with room to grow
Substantial tools and ongoing systems live here. These cards are generated from the project collection, so new work starts as a Markdown or MDX record.
project-music-creatorin development
Projectgenerative toolMusic Creator Harness
An experimental harness for exploring model-assisted music creation, iterative generation, and review workflows.
project-synthetic-datain development
Projectai toolSynthetic Data Designer
A tool for turning a natural-language dataset specification into structured synthetic data for workflows such as supervised fine-tuning.
03 / Lab notebook
Experiments & tests
Benchmarks, failed attempts, autonomous runs, model-training observations, and visual trials belong here—even when they are not polished products.
exp-qwen-blender-cityplanned
Experimentwiggum loopCan Qwen Build a City Autonomously?
A planned field note for an autonomous Blender run coordinated through a long-running Wiggum loop.
exp-gradient-accumulation-sftplanned
ExperimenttrainingTesting Gradient Accumulation During SFT
A structured place to compare training behavior while changing accumulation settings during supervised fine-tuning.
exp-visual-novel-planningplanned
Experimentmini projectAutonomous Visual Novel Planning
A future notebook entry for evaluating agent-assisted planning and review in a Ren'Py-oriented workflow.
04 / Model work
Training, not just inference
The model section is designed around the engineering process: tokenizer and dataset work, architecture experiments, training progression, outputs, failures, and lessons learned.
model-diffusion-v1research
ModelDiffusion ModelSelf-Trained Diffusion Study
A diffusion-model research track focused on dataset work, training configuration, checkpoint progression, generated outputs, and evaluation.
model-yunagptresearch
ModelLanguage ModelYunaGPT
A self-trained decoder language-model project covering architecture experiments, tokenizer work, data preparation, pretraining, instruction tuning, and alignment studies.