# AI Helpers

> A BSD-3-Clause suite of 15 Python libraries for AI development, organized into groups:
> Core (OS), Audio & voice (audio, vocal diarization/transcription),
> Video & capture, Media acquisition (YouTube, podcast), Storage & transfer (S3, SFTP),
> and Misc (Markdown → DOCX/PPTX/PDF, content-addressed caching, PCA positioning maps,
> hardware-aware local LLM/VLM selection over Ollama, and approximate-nearest-neighbour
> vector-search routing).
> By Warith Harchaoui, Mohamed Chelali and Bachir Zerroug. Online since April 2024.

Home: https://harchaoui.org/warith/ai-helpers/
Releases (Atom feed): https://harchaoui.org/warith/ai-helpers/releases.xml
Install: every helper is its own package on PyPI — install by name, e.g. `pip install os-helper audio-helper vocal-helper youtube-helper md2star`.

## Libraries

### Core
- [OS Helper](https://github.com/warith-harchaoui/os-helper): utility functions for cross-platform OS operations, file handling, hashing, config, timing and system tasks. The shared foundation every other helper builds on.

### Audio & voice
- [Audio Helper](https://github.com/warith-harchaoui/audio-helper): load, convert, source-separate (optional Demucs), split and concatenate audio files. Also in Rust: [audio-helper-rs](https://github.com/warith-harchaoui/audio-helper-rs), file-level audio utilities on ffmpeg/ffprobe (single static binary, cargo install, not on PyPI, a smaller v0.1 surface).
- [Vocal Helper](https://github.com/warith-harchaoui/vocal-helper): async pipeline turning audio into diarized, transcribed utterances plus an optional rolling LLM summary; online-streaming and offline-batch modes on Silero VAD, whisper.cpp, pyannote/NeMo and Ollama, with CLI/Python/HTTP/MCP interfaces.

### Video & capture
- [Video Helper](https://github.com/warith-harchaoui/video-helper): multi-backend frame extraction (VidGear/PyAV/ffmpeg-pipe), load, convert and work with subtitle formats for video files.
- [Capture Helper](https://github.com/warith-harchaoui/capture-helper): live multi-source capture layer. Camera and microphone iterators (composing with the Video Helper and Podcast Helper contracts) plus a browser scene configurator to design and preview multi-source scenes. Also in Rust: [capture-helper-rs](https://github.com/warith-harchaoui/capture-helper-rs), live microphone capture via cpal, no ffmpeg needed (cargo install, not on PyPI, a smaller v0.1 surface).

### Media acquisition
- [YouTube Helper](https://github.com/warith-harchaoui/youtube-helper): download video/audio/thumbnails from YouTube, Vimeo, DailyMotion, SoundCloud and Twitch via yt-dlp, plus ffmpeg post-processing and no-API engagement metadata. Also in Rust: [youtube-helper-rs](https://github.com/warith-harchaoui/youtube-helper-rs), a typed wrapper around the yt-dlp binary (cargo install, not on PyPI, a smaller v0.1 surface).
- [Podcast Helper](https://github.com/warith-harchaoui/podcast-helper): universal audio-stream consumption (URL-in to PCM-out) from files, direct URLs, RSS/Atom feeds and any yt-dlp source, with Shannon-correct resampling. Also in Rust: [podcast-helper-rs](https://github.com/warith-harchaoui/podcast-helper-rs), URL-in/PCM-out, delegating yt-dlp sources to youtube-helper-rs (cargo install, not on PyPI, a smaller v0.1 surface).

### Storage & transfer
- [Bucket Helper](https://github.com/warith-harchaoui/bucket-helper): AWS S3 and S3-compatible object storage (MinIO, Backblaze B2, DigitalOcean Spaces, Cloudflare R2, Wasabi) via boto3.
- [SFTP Helper](https://github.com/warith-harchaoui/sftp-helper): interact with SFTP servers via the system OpenSSH `sftp` client, with strict host-key verification and a remote_tempfile helper, using SSH key credentials.

### Misc
- [md2star](https://github.com/warith-harchaoui/md2star): Markdown to DOCX/PPTX/PDF bridge built on Pandoc, with curated branded templates, Mermaid diagram rendering and bibliography support. A `twin` command runs the reverse path: it reconstructs an editable Markdown file plus an `assets/` folder from a PDF (or anything LibreOffice converts to one), and can re-author node-and-edge figures as Mermaid. A local browser GUI (`md2star gui`) gives a live PDF preview. PyPI: `pip install md2star`. Also in Rust: [md2star-rs](https://github.com/warith-harchaoui/md2star-rs), a pure-Rust Markdown → DOCX/PPTX writer, no Pandoc, single static binary (cargo install, not on PyPI, a smaller v0.1 surface).
- [Wallet Helper](https://github.com/warith-harchaoui/wallet-helper): never run the same heavy call twice. Persistent, content-addressed memoization (keys on a file's content or bytes, not just arguments) plus single-flight so concurrent identical calls collapse into one, in-process and, via a fenced SQLite lease or an optional HTTP dedup server, across processes and hosts. A local-first toolbox, close in spirit to OS Helper.
- [Standpoint](https://github.com/warith-harchaoui/standpoint): turn a comparison table (options as rows, criteria as columns, numbers in the cells) into a labelled 2D positioning map, a short written analysis and a YAML of coordinates, in one command. Ordinary PCA plus the hand-work: orient around a reference option, name the axes in plain words in the table's own language, colour and label the points. Local-first (figure hand-authored as SVG and rasterised via resvg; axis names come from a local Ollama model routed through Best Engine AI Helper). PyPI: `pip install standpoint`.
- [Best Engine AI Helper](https://github.com/warith-harchaoui/best-engine-ai-helper): pick and pull the best local LLM/VLM for the current machine. It detects available memory (Apple Silicon unified pool, NVIDIA VRAM, or system RAM), consults a bundled model catalog, selects the highest-scoring model that fits a configurable safety headroom, pulls it via Ollama, and writes an env file downstream projects source. Ranking prefers models that can emit valid structured JSON when a task needs it. Cheap `text_model()` / `vision_model()` resolvers answer "which model do I use?" without re-probing hardware. `catalog update` and `hardware update` refresh the cached model and machine facts, and a bilingual browser GUI (`best-engine-ai-helper gui`) shows the hardware snapshot and the recommendation. Local-first (models run through Ollama). PyPI: `pip install best-engine-ai-helper`.
- [ANN Router](https://github.com/warith-harchaoui/ann-router): the vector-search sibling of Best Engine AI Helper — same philosophy, measure the criteria, select the engine, return a discussable rationale, applied to approximate-nearest-neighbour search. Describe an ANN problem in measured terms (dataset size, dimensionality, recall target, update pattern, memory budget) and it selects, justifies, and can instantiate the right backend among exact (brute force), turbovec, HNSW (hnswlib), FAISS (IVF/PQ), Annoy, Qdrant and pgvector. Dependency-free at import time: each engine's optional dependency loads lazily, and a backend whose dependency is absent simply reports itself unavailable while the router routes around it. An indispensable component for RAG. PyPI: `pip install ann-router`.
- [Elbow Helper](https://github.com/warith-harchaoui/elbow-helper): noise-robust knee/elbow detection that reports a knee with uncertainty or explicitly abstains, rather than always returning a point estimate on a straight line or on pure noise. Wraps a from-scratch, NumPy-only knee locator in a conservative decision procedure — strong, unique, persistent, reproducible, and unlikely under a no-knee model — prioritising fewer false positives over never abstaining. Also exposes `robust_knees` for the multi-knee case. Dependency-free beyond `numpy` and `os-helper`, including its hand-authored SVG diagnostic figure. PyPI: `pip install elbow-helper`.

The six above are bundled in the `misc` extra. Five of the fifteen also have
an independent Rust rewrite, noted inline above (separate GitHub repos,
install via `cargo`, not on PyPI, a smaller v0.1 surface, not a drop-in
replacement).

## License

BSD-3-Clause for the 15 bundled helpers (same as scikit-learn / numpy / scipy).