Open-source AI demos you can understand, run, and remix.
The shortest path from “that demo looks impossible” to “I built it.”
Explore the demos · Run one now · Browse the catalog · Contribute
The demos you see on X and LinkedIn are a whole production: live models, carefully chosen inputs, polished interfaces, recording, editing, and timing. This repository strips each one down to the useful core so you can learn the idea, run it yourself, and turn it into something new.
No mystery framework. No provider lock-in. No private production machinery. Just small, readable builds with one clear result.
| Demo | Model(s) | What you build | Input → output | Run it |
|---|---|---|---|---|
| Draw object boxes with a vision model | Qwen3.8-27B | Ask a vision model for normalized coordinates and render the boxes | Photo + labels → annotated image | Local or hosted |
| Read form checkboxes | Muse Glimmer 30B | Read checkbox labels and states, then annotate the form | Form image → annotated image | Local or hosted |
There is one row per demo. When another model is verified with the same code, its name joins the existing Model(s) cell instead of creating a duplicate row.
More demos will land across vision, audio, documents, agents, video, and text-and-code. The complete, growing index lives in the demo catalog.
Clone the repository and launch the first demo:
git clone https://github.com/maziyarpanahi/ai-demos.git
cd ai-demos/vision/draw-object-boxes
python3 -m venv .venv
source .venv/bin/activate
python -m pip install -r requirements.txt
export OPENAI_BASE_URL="http://localhost:11434/v1"
export OPENAI_API_KEY="local"
export MODEL="your-vision-model-id"
python detect.py your-photo.jpg --labels "person, car, bicycle"Use a local OpenAI-compatible server or point the same demo at a hosted endpoint. Each demo tells you exactly what it needs and what leaves your machine.
- Small: one focused idea, not an application framework.
- Readable: one obvious entry point and code you can understand quickly.
- Reproducible: copy-paste setup, explicit requirements, and expected output.
- Portable: standard interfaces over provider-specific SDKs whenever practical.
- Honest: clear local-versus-hosted boundaries, privacy notes, and known limitations.
- Hackable: the important prompt, schema, and rendering logic stay easy to find and change.
<category>/<outcome>/
├── README.md # what it does, how to run it, how to adapt it
├── <entry-point> # the smallest readable implementation
├── requirements.txt # direct dependencies
└── assets/ # optional, licensed example inputs and outputs
Categories live directly at the repository root, so the collection stays visible without turning the root into hundreds of folders. Demo paths are named after outcomes rather than providers, which keeps links useful as models change.
These are not frame-for-frame replicas of the original posts. Social versions may include custom interfaces, capture tooling, larger pipelines, and private operational details. The repository version preserves the technique and the learning value while removing everything you do not need to reproduce the core result.
Start with a demo, swap the input, change the prompt, try another compatible model, or combine the idea with your own product. If you improve one, add a new runtime, or build a clean variation, read CONTRIBUTING.md.
If a demo helps you make something, star the repository so you can find it again—and share your remix. That is how this collection gets better.
Code in this repository is released under the MIT License.