Stillk

Stillk is an open-source scaffolding CLI for machine learning, deep learning, LLM, RAG, and AI agent projects. It follows a create-react-app-inspired workflow so teams can generate a production-ready project skeleton in seconds.

Why Stillk

• Generate a complete starter project with real folders and working files
• Support ML, DL, LLM, RAG, and AI agents from day one
• Modular registry-based template system for future growth
• Open source and designed for community-driven extension

Features

• stillk init <project-name> to bootstrap a project
• Interactive project selection when flags are omitted
• Built-in templates for ML and LLM flows
• Pre-generated pyproject.toml, README.md, Dockerfile, .env.example, tests, and source folders
• Future plugin-ready command: stillk add mlflow, stillk add rag, stillk add fastapi

Installation

pip install stillk

For local development:

python -m venv .venv
source .venv/bin/activate
pip install -e .[dev]

Quick start

stillk init meu-projeto --project-type ml --framework scikit-learn
stillk init meu-llm --project-type llm --framework transformers
stillk init

AI Engineering Toolkit commands

Stillk now includes runtime tooling to help during development, training and maintenance:

• stillk inspect — analyze the current project for frameworks, models, datasets, Docker, GPU and give warnings.
• stillk doctor — run quick environment checks (Python, pip, Docker, GPU, PyTorch/TensorFlow).
• stillk clean — detect caches and temporary artifacts and optionally remove them (--yes).
• stillk train — detect training scripts and (optionally) execute them (--execute).
• stillk eval — run or detect evaluation pipelines and report metrics.
• stillk benchmark — lightweight benchmarking scaffold for models and pipelines.
• stillk run <component> — run project components such as api, training, inference.

Analysis and context

• stillk context — generate LLM-optimized project context (architecture, entrypoints, important files). Use --output to write a Markdown summary.
• stillk explain <target> — explain a file, module, or the whole project (stillk explain project).
• stillk diff — analyze git changes and report technical impact on models, prompts, deps, configs, and pipelines.
• stillk tokens — estimate token usage for files/prompts/context (--text, --price-per-1k).
• stillk optimize — suggest optimizations for code, prompts, RAG, cache, and deps. Apply safe fixes only with --apply --yes.
• stillk deps — find unused, heavy, outdated, or conflicting dependencies.

Data, models, RAG, and prompts

• stillk dataset inspect <file> — inspect CSV/JSON/JSONL/Parquet datasets (rows, dtypes, missing, duplicates, class distribution). Example: stillk dataset inspect data/train.csv --target label.
• stillk model — inspect model configs and checkpoints (framework, size, parameters when available).
• stillk rag inspect — inspect RAG directories, chunking, top-k, embeddings, and vector store settings.
• stillk rag eval — structural RAG evaluation / readiness checks (--dataset optional).
• stillk prompt list — list prompt files and versions.
• stillk prompt test — validate prompts against an eval dataset when configured.

Environment, profiling, and experiments

• stillk env — report Python, OS, GPU, CUDA, Docker, and key dependencies. Use --report to write stillk-env-report.md.
• stillk profile — profile a script or command for wall time, CPU/GPU, and memory (--script / --command).
• stillk reproduce — collect commit, Python, deps, dataset, model, hardware, CUDA, and seeds into reproduce.json.
• stillk experiment start — register a local experiment (model, dataset, config, metrics, commit, timestamp).
• stillk experiment list — list experiments stored under .stillk/experiments/.

The interactive mode asks for:

• project type
• framework
• output folder
• project metadata

Generated project structure

meu-projeto/
├── src/
│   ├── data/
│   ├── features/
│   ├── models/
│   ├── training/
│   ├── inference/
│   ├── api/
│   ├── agents/
│   ├── graph/
│   ├── prompts/
│   ├── retrieval/
│   ├── embeddings/
│   ├── vectorstore/
│   └── evals/
├── data/
│   ├── raw/
│   ├── processed/
│   └── external/
├── configs/
├── notebooks/
├── tests/
├── scripts/
├── Dockerfile
├── pyproject.toml
├── README.md
├── .env.example
├── .gitignore
├── main.py
├── preprocessing.py
└── .github/

Supported types

• Machine Learning
• Deep Learning
• LLM
• RAG
• AI Agents

Supported frameworks

• Scikit-learn
• PyTorch
• XGBoost
• Transformers
• LangGraph

Publishing to PyPI

To build the package:

python -m pip install -U build twine
python -m build

To publish:

python -m twine upload dist/*

Before publishing, configure your PyPI API token in your environment or use a trusted publishing workflow.

Contributing

Contributions are welcome. Please open an issue or pull request with tests for new templates or features.

License

MIT

Copyright (c) 2026 Lucas Balduino - Stillk