ti-edgeai-skills
Agent skills for the Texas Instruments Edge AI SDK. They cover training, compiling, deploying, running and profiling neural networks on TI processors with a C7x DSP and MMA accelerator (TDA4VM, AM68A, AM69A, AM67A, AM62A), using TI's edgeai-tidl-tools on the PC and the Edge AI Linux SDK with edgeai-gst-apps on the board.
The skills work with Claude Code and Codex. They are written against the SDK rather than any particular model or board.
Skills
| Skill | Purpose |
|---|---|
ti-edgeai-dev | SDK components, SoC lookup table, tools-to-SDK version matching, board services, troubleshooting |
ti-edgeai-train-model | Architecture and toolchain choice for TIDL, datasets and group-wise splits, deployment-matched evaluation |
ti-edgeai-import-model | ONNX preflight, TIDL compile in Docker, accuracy checks, packaging, deployment, host-versus-board comparison |
ti-edgeai-generate-config | Generate, validate and run edgeai-gst-apps configs (camera, video, images, display, stream, mosaic) |
ti-edgeai-profile-pipeline | Model latency, per-layer C7x cycles, GStreamer element latency, optimization levers |
A typical project uses them in this order: dev for orientation, train-model, import-model, generate-config, profile-pipeline.
Each skill is a folder with a SKILL.md, plus references/, scripts/, assets/, tests/ and evals/ where needed.
Installation
Claude Code plugin
In Claude Code:
/plugin marketplace add saeedahmadicp/ti-edgeai-skills
/plugin install ti-edgeai-skills@ti-edgeai-skills
To work from a local clone instead, pass its path to /plugin marketplace add.
Skills directory (Claude Code or Codex)
git clone https://github.com/saeedahmadicp/ti-edgeai-skills.git
cd ti-edgeai-skills
scripts/install.sh
scripts/install.sh symlinks the skills into the folder the agent reads. It does not delete or overwrite anything.
| Command | Folder |
|---|---|
scripts/install.sh | ~/.claude/skills |
scripts/install.sh --agent claude --scope repo | <your-project>/.claude/skills |
scripts/install.sh --agent codex | ~/.agents/skills |
scripts/install.sh --agent codex --scope repo | <your-project>/.agents/skills |
With --scope repo the folder is the git root of the current directory, so run the script from inside the project that should
get the skills (for example /path/to/ti-edgeai-skills/scripts/install.sh --scope repo). Add --dry-run to preview.
A root plugin.json (portable plugin layout) and per-skill agents/openai.yaml are included for Codex. Codex discovery has not been tested; the paths follow OpenAI's documentation.
Usage
Describe the task and the matching skill loads. In Codex a skill can also be named explicitly with $.
Which edgeai-tidl-tools release do I need for my TI board?
$ti-edgeai-import-model compile model.onnx for my board and deploy it
Run my detector on the USB camera and save the output
Why is my pipeline slower than 30 fps?
Requirements
- A TI Edge AI board reachable over SSH, running the Edge AI SDK Linux image.
- Docker on a PC, for the TIDL tools container.
- Python 3 with PyYAML for the config tools. numpy, OpenCV and onnx are needed by some scripts and are present in the container.
- A GPU only for training.
Board addresses are placeholders (root@<board-ip>). Nothing in the repository stores credentials.
Supported hardware
| SoC | Status |
|---|---|
| TDA4VM (J721E, AM68PA) | Tested end to end on Edge AI SDK 11.0 |
| AM68A, AM69A, AM67A, AM62A | Values taken from TI documentation, not tested |
SoC-specific values (tools SOC, gst-apps SOC, quantization support) are collected in
platforms.md. What has and has not been run is listed in
docs/verification-status.md.
Documentation
- docs/architecture.md: how the skills fit together and the design rules
- docs/testing.md: test levels, hardware verification protocol, running the evals
- docs/verification-status.md: what was run, on which hardware
- CHANGELOG.md
Testing
pip install pyyaml numpy opencv-python-headless onnx
tests/run_tests.sh
This runs the lint and the unit tests. No board or GPU is required. Tests that need a missing dependency are skipped.
Contributing
See CONTRIBUTING.md. Reports from boards other than the TDA4VM are the most useful contribution; use the issue templates.
License
MIT, see LICENSE. TI's tools, SDK and model zoo are under their own licenses and are not redistributed here.