AIE-Skills

Bộ Kiro Skills & Steering cho workflow AI/ML Engineering — từ setup Python project, fine-tune LLM, đến deploy inference server.

---

Quickstart

One-liner Install

# Install vào project hiện tại
curl -fsSL https://raw.githubusercontent.com/jayll1303/AIEKit/main/install.sh | bash

# Install vào thư mục cụ thể
curl -fsSL https://raw.githubusercontent.com/jayll1303/AIEKit/main/install.sh | bash -s -- /path/to/project

# Install globally (vào ~/.kiro/)
curl -fsSL https://raw.githubusercontent.com/jayll1303/AIEKit/main/install.sh | bash -s -- --global

Script chỉ copy components chưa tồn tại — không overwrite file đã có.

Smart Install (recommended)

Dùng skill aie-skills-installer trong Kiro — nó sẽ:

1. Scan codebase target (deps, imports, Dockerfiles, notebooks...)
2. Recommend chỉ skills có signal cụ thể từ project
3. Chờ user confirm trước khi cài
4. Cài selective + steering files tương ứng

Manual Install

git clone https://github.com/jayll1303/AIEKit.git /tmp/aie-skills
bash /tmp/aie-skills/.kiro/install.sh
rm -rf /tmp/aie-skills

---

Skills (27)

Skill	Mô tả
aie-skills-installer	Analyze target project codebase và đề xuất chỉ cài skills cần thiết (tránh cài toàn bộ tốn context)
arxiv-reader	Đọc và phân tích paper arXiv qua HTML
docker-gpu-setup	Dockerfile & docker-compose cho GPU/CUDA workloads
experiment-tracking	Selfhosted experiment tracking với MLflow / W&B
fastapi-at-scale	Build production-grade FastAPI at scale: project structure, async SQLAlchemy, Alembic migrations, JWT auth, rate limiting, testing với httpx, deploy uvicorn/gunicorn/Docker
freqtrade	Phát triển crypto trading strategies với Freqtrade
hf-hub-datasets	Download, upload, stream models & datasets từ HuggingFace Hub
hf-speech-to-speech-pipeline	Architecture patterns cho huggingface/speech-to-speech queue-chained pipeline: STT/LLM/TTS handlers, VAD, progressive streaming
hf-transformers-trainer	Fine-tune & align LLMs với Trainer, TRL, PEFT (LoRA/QLoRA)
k2-training-pipeline	Train speech models với Next-gen Kaldi: k2 (FSA/FST loss), icefall (Zipformer/Conformer recipes), lhotse (data prep)
llama-cpp-inference	Chạy GGUF models locally với llama-server, llama-cli, llama-cpp-python (CPU+GPU)
model-quantization	Quantize LLMs với GGUF, GPTQ, AWQ, bitsandbytes
notebook-workflows	Tạo & chỉnh sửa Jupyter/Colab notebooks programmatically
ollama-local-llm	Chạy và quản lý local LLMs với Ollama: pull, run, Modelfile, REST API
paddleocr	OCR với PaddlePaddle: text detection, recognition, fine-tuning, dataset prep, PP-OCRv5, PP-StructureV3
python-ml-deps	Cài ML deps với uv, xử lý CUDA version conflicts
python-project-setup	Bootstrap Python projects với uv, ruff, pytest
python-quality-testing	Type annotations, Hypothesis testing, mutation testing
sglang-serving	Serve LLMs với SGLang: RadixAttention prefix caching, structured output (JSON/regex/EBNF)
sherpa-onnx	Offline speech processing: ASR, TTS, VAD, speaker diarization, speech enhancement
tensorrt-llm	Optimize LLM inference với NVIDIA TensorRT-LLM: engine building, FP8/INT4, kernel fusion
text-embeddings-inference	Deploy embedding/reranker models với HuggingFace TEI
text-embeddings-rag	RAG pipelines với sentence-transformers, FAISS, ChromaDB, Qdrant
triton-deployment	Deploy models trên NVIDIA Triton Inference Server
ultralytics-yolo	Train, predict, export, deploy YOLO models (detect, segment, classify, pose, OBB) với Ultralytics
unsloth-training	Fine-tune LLMs 2x faster, 70% less VRAM với Unsloth: SFT/DPO/GRPO, export GGUF/vLLM
vllm-tgi-inference	Serve LLMs locally với vLLM hoặc TGI

Steering (6)

File	Inclusion	Mô tả
kiro-component-creation.md	always → auto khi install	Quy tắc tạo Steering, Skills, Hooks, Powers cho Kiro
notebook-conventions.md	fileMatch (**/*.ipynb)	Conventions khi làm việc với file .ipynb
ml-training-workflow.md	auto	Conventions cho ML training & fine-tuning workflows
inference-deployment.md	auto	Conventions cho model serving & deployment
python-project-conventions.md	auto	Conventions cho Python projects: uv, ruff, pytest, CUDA deps
gpu-environment.md	fileMatch (Dockerfile*, docker-compose*)	Conventions cho GPU Docker containers

Hooks (6)

Hook	Event	Mô tả
update-readme-index	fileEdited	Auto-update README index khi edit component, commit + push (cần confirm)
readme-index-on-create	fileCreated	Auto-update README index khi tạo component mới, commit + push (cần confirm)
readme-index-on-delete	fileDeleted	Auto-update README index khi xóa component, commit + push (cần confirm)
skill-quality-gate	fileCreated	Check SKILL.md mới theo best practices + update interconnection map
skill-quality-on-edit	fileEdited	Check SKILL.md đã sửa theo best practices + interconnection map
steering-consistency	fileCreated	Check steering mới: frontmatter, domain overlap, cross-references

Powers (2)

Power	MCP Server	Mô tả
power-huggingface	HF MCP Server (remote HTTP)	Search models, datasets, papers, spaces trên HuggingFace Hub. Compare models, check configs, discover trending papers
power-gpu-monitor	mcp-system-monitor (local Python)	Monitor GPU/VRAM/CPU real-time, estimate memory cho ML models, diagnose OOM errors

Mỗi power bao gồm: POWER.md + mcp.json + steering/ workflows.