AIEKit

Modal job orchestration with .map, .starmap, .spawn, .spawn_map, @modal.batched. Use when fanning out work, polling job IDs, capping concurrency, or batching GPU calls.

Use this skill for Modal Sandbox lifecycle and interactive isolated execution. Trigger when the user needs to run untrusted or user-supplied code in a sandbox, keep a long-lived controller process alive, stream stdout from `exec()` calls, reattach to an existing sandbox, upload files at runtime, expose a tunneled HTTP or TCP service, or checkpoint and restore sandbox state. Do not use it for regular `@app.function` deployments, LLM serving, or training workflows.

Analyze project and recommend relevant AIE-Skills. Use when installing, setting up ML skills, bootstrapping AI/ML workflow, or sharing skills across repos.

Read and analyze arXiv papers via HTML. Use when user mentions arxiv, paper, research paper, academic paper, arxiv ID, or asks to read/summarize/analyze a scientific paper.

Diagnose and clean disk space on Linux servers, especially ML/LLM servers with Docker. Use when disk full, df shows high usage, du mismatch, Docker eating space, deleted files still consuming space, or need to free up disk on EC2/GPU server.

Build reproducible GPU-enabled Docker containers for ML/CUDA workloads. Use when writing GPU Dockerfiles, selecting NGC base images, configuring docker-compose GPU passthrough, setting up NVIDIA Container Toolkit, or installing uv inside Docker containers.

Set up selfhosted experiment tracking with MLflow or W&B. Use when deploying MLflow servers, configuring W&B selfhosted, logging metrics and artifacts, comparing experiment runs, managing model registry, or troubleshooting offline sync.

Build production-grade FastAPI at scale. Use when structuring projects, async SQLAlchemy, Alembic migrations, JWT auth, rate limiting, testing with httpx, deploying uvicorn/gunicorn/Docker.

Develop crypto trading strategies with Freqtrade. Use when creating IStrategy classes, writing populate_indicators/entry/exit, running backtesting/hyperopt, configuring bot JSON, implementing callbacks (custom_stoploss, DCA, leverage), or downloading OHLCV data.

Download, upload, and stream HuggingFace models and datasets. Use when downloading pretrained models with snapshot_download, loading datasets with load_dataset, pushing to Hub with push_to_hub, streaming large datasets, or managing private repos and gated models.

Best practices and architecture patterns for huggingface/speech-to-speech queue-chained pipeline. Use when building speech-to-speech pipelines, adding STT/LLM/TTS handlers, wiring queue-based audio processing stages, implementing VAD with progressive streaming, or designing real-time voice agent architectures.

Fine-tune and align LLMs locally with HuggingFace Trainer, TRL, PEFT. Use when configuring TrainingArguments, applying LoRA or QLoRA, running SFT/DPO/GRPO alignment, estimating VRAM requirements, debugging OOM errors, or setting up gradient checkpointing.

Train speech models with Next-gen Kaldi ecosystem: k2 (FSA/FST loss), icefall (Zipformer/Conformer recipes), lhotse (data prep). Use when training ASR/TTS models, preparing speech datasets, exporting to ONNX for sherpa-onnx, or fine-tuning pre-trained speech models.

Run GGUF models locally with llama.cpp. Use when launching llama-server, running llama-cli, using llama-cpp-python, configuring GGUF inference on CPU+GPU, or tuning continuous batching.

Brainstorm ML/AI decisions: training strategy, model selection, serving engine, quantization, pipeline architecture. Use when comparing approaches, evaluating tradeoffs, or planning ML workflows

Quantize large language models to reduce VRAM and accelerate inference with GGUF, GPTQ, AWQ, bitsandbytes. Use when converting models to GGUF format, applying GPTQ/AWQ weight-only quantization, loading models in 4-bit or 8-bit with BitsAndBytesConfig, or comparing quantization methods for VRAM budget.

Create, edit, and manage Jupyter/Colab .ipynb notebooks programmatically. Use when creating notebooks, adding/editing cells, parsing notebook output, handling Colab features (GPU, Drive mount), or converting notebook formats.

Run and manage local LLMs with Ollama. Use when running ollama run, ollama pull, creating Modelfiles, serving local LLM APIs, using ollama serve, or integrating Ollama REST/OpenAI-compatible API.

Build OpenAI-compatible audio/speech APIs with FastAPI, dynamic batching, and streaming synthesis. Use when building TTS servers, audio inference APIs, or wrapping audio ML models behind HTTP.

Instrument apps with OpenTelemetry for distributed tracing, metrics, and logs. Use when setting up OTel SDK, auto-instrumenting Python/Node.js, configuring Collector pipelines, or choosing sampling.

OCR with PaddlePaddle: PaddleOCR text detection, recognition, fine-tuning, dataset prep, PP-OCRv5, PP-StructureV3. Use when training OCR models, preparing OCR datasets, fine-tuning detection/recognition, exporting PaddleOCR inference models, or deploying OCR pipelines.

Install ML/AI Python deps with uv, resolve NVIDIA version conflicts. Use when installing PyTorch with CUDA index URLs, building Flash-Attention from source, resolving CUDA/cuDNN/driver version conflicts, installing bitsandbytes, or configuring pyproject.toml with CUDA-specific extras.

Bootstrap and configure Python projects with uv, ruff, and pytest. Use when running uv init, setting up pyproject.toml, configuring ruff linting and formatting, configuring pytest, adding pre-commit hooks, or migrating from pip or Poetry.

Strengthen Python code quality and test suites. Use when adding type annotations, writing contract docstrings, auditing exception handling, implementing Hypothesis property-based testing, running mutmut mutation testing, or performing code quality reviews.

Build superfast AI decision layers with semantic-router. ALWAYS use for Route, SemanticRouter, HybridRouter, dynamic routes with function calling, intent classification, Pinecone/Qdrant index.

Serve LLMs with SGLang for structured generation and high-throughput inference. Use when launching SGLang server, using RadixAttention, constrained JSON/regex output, or structured decoding.

Offline speech processing with sherpa-onnx: ASR (streaming/non-streaming), TTS (Piper/Kokoro/Matcha/VITS), VAD, speaker diarization, speech enhancement. Use when running speech-to-text, text-to-speech, speaker identification, voice activity detection, or audio processing locally without internet.

Optimize LLM inference with NVIDIA TensorRT-LLM. Use when running trtllm-build, converting HF checkpoints, serving with FP8/INT4, or maximizing GPU throughput.

Deploy and serve embedding/reranker models with HuggingFace TEI. Use when launching TEI Docker, configuring embeddings API, choosing embedding models, tuning batch performance, or building RAG pipelines with TEI.

Build local RAG pipelines with sentence-transformers, FAISS, ChromaDB, Qdrant. Use when generating embeddings, setting up semantic search, or optimizing retrieval with re-ranking and hybrid search.

Deploy ML models on NVIDIA Triton Inference Server. Use when writing config.pbtxt, structuring model_repository, building ensemble pipelines, configuring dynamic batching, writing tritonclient code, or debugging Triton model loading failures.

Train, predict, export, and deploy YOLO models with Ultralytics. Use when running yolo train/predict/val/export, building custom object detection/segmentation/classification/pose/OBB pipelines, preparing data.yaml datasets, or deploying YOLO to ONNX/TensorRT/CoreML/edge devices.

Fine-tune LLMs 2x faster with 70% less VRAM using Unsloth. Use when using FastLanguageModel, Unsloth SFT/DPO/GRPO, 2x faster fine-tuning, or exporting to GGUF/vLLM.

Deploy and serve LLMs locally with vLLM or TGI. Use when launching vllm serve, running TGI Docker, configuring tensor parallelism, serving quantized models, using OpenAI-compatible API, tuning KV cache, or diagnosing OOM errors.

Analyze project and recommend relevant AIE-Skills. Use when installing, setting up ML skills, bootstrapping AI/ML workflow, or sharing skills across repos.

Diagnose and clean disk space on Linux servers, especially ML/LLM servers with Docker. Use when disk full, df shows high usage, du mismatch, Docker eating space, deleted files still consuming space, or need to free up disk on EC2/GPU server.

Build superfast AI decision layers with semantic-router. ALWAYS use for Route, SemanticRouter, HybridRouter, dynamic routes with function calling, intent classification, Pinecone/Qdrant index.

Brainstorm ML/AI decisions: training strategy, model selection, serving engine, quantization, pipeline architecture. Use when comparing approaches, evaluating tradeoffs, or planning ML workflows

Read and analyze arXiv papers via HTML. Use when user mentions arxiv, paper, research paper, academic paper, arxiv ID, or asks to read/summarize/analyze a scientific paper.

Develop crypto trading strategies with Freqtrade. Use when creating IStrategy classes, writing populate_indicators/entry/exit, running backtesting/hyperopt, configuring bot JSON, implementing callbacks (custom_stoploss, DCA, leverage), or downloading OHLCV data.