Runnable, real-world AIMU programs, larger than the notebook cells, organized by theme. Each subdirectory has its own README with run commands and a link to the relevant how-to guide.
| Example | What it shows | Start with |
|---|---|---|
| text-refinement/ | A generate, judge, refine loop over text, implemented four ways (code loop, Agent, EvaluatorOptimizer, simulated annealing). GPU-free, Ollama-only. |
python examples/text-refinement/epic_loop.py |
| image-refinement/ | The same loop over images (diffusion + vision evaluator), five variants including img2img. Needs the [hf] extra and a GPU. |
python examples/image-refinement/hotdog_loop.py |
| news-summarizer/ | One task (summarize recent AI news) solved with Agent, Chain, Parallel, and OrchestratorAgent, selected via --method. |
python examples/news-summarizer/news_summarizer.py --method agent |
| personal-assistant/ | A single-user, always-on assistant (OpenClaw / Hermes style) from AIMU primitives: a Channel (CLI or a WebSocket web_assistant.py), a Scheduler for proactive messages, and a SkillAgent that authors its own skills and scripts. |
python examples/personal-assistant/assistant.py |
| web/ | Streamlit and Gradio chat apps (showcase + full-featured). Needs pip install aimu[web]. |
streamlit run examples/web/streamlit_chatbot.py |
| skills/ | Demo SKILL.md skills (haiku-poet, unit-converter) for SkillAgent discovery, exposed as aimu.paths.skills. |
08 - Agent Skills |
The text-refinement and image-refinement families are deliberately the same task in two modalities. Read them side by side to see that the diff between "who drives the loop" (code vs. agent vs. workflow) and "which search strategy" (greedy vs. hill-climbing vs. annealing) is independent of whether the artifact is a string or an image.
The example test suites are kept out of the default pytest run. Run them explicitly:
pytest examples/ # all example suites
pytest examples/text-refinement/tests # one suitenotebooks/: interactive, subsystem-by-subsystem demos (numbered 01-24)- How-to guides: task-oriented recipes