Skills for your AI
Deploys, tunes, and troubleshoots vLLM inference servers for production APIs, offline batch inference, quantized serving, and performance issues. Use when the user needs vLLM launch commands, quantization advice, throughput or TTFT fixes, or help choosing betw
Generates and configures Modal serverless GPU code for ML training, inference, and batch workloads, including GPU selection, web endpoints, volumes, secrets, and scheduling. Use when the user wants to run Python workloads on Modal, pick a GPU, serve a model as
Designs production LLM architectures covering serving, fine-tuning, RAG, and multi-model orchestration with measurable latency, throughput, and cost targets. Use when the user needs an LLM serving stack, quantization choice, fine-tuning pipeline, RAG retrieval
Produces validated LoRA/QLoRA supervised fine-tuning adapter configurations with target modules, rank, alpha, learning rate, batch size, and Unsloth defaults. Use when configuring a LoRA/QLoRA adapter, choosing rank or learning rate, checking effective batch s
Deploys and optimizes ML models for production inference, covering serving infrastructure, quantization, monitoring, edge and batch serving, and auto-scaling. Use when the user needs to deploy a model, cut inference latency or cost, design serving infrastructu
Guides data analysts through machine learning projects from preprocessing and feature engineering to model selection, tuning, evaluation, interpretability, deployment, and applied use cases. Use when the user asks for help cleaning data for ML, choosing or tun
Implements, fine-tunes, pretrains, quantizes, and deploys LLMs with LitGPT's pretrained architectures and LoRA/QLoRA workflows. Use when the user wants to load a LitGPT model for inference, fine-tune on their own dataset, pretrain from scratch, prepare tokeniz
Explains the RWKV architecture and guides installation, inference, Transformer comparisons, and fine-tuning. Use when the user asks how RWKV works, wants to install or set up RWKV, run or stream inference, compare RWKV with Transformers, or fine-tune an RWKV m
Converts molecular SMILES strings into numerical feature vectors using molfeat featurizers (fingerprints, 2D descriptors, pretrained embeddings, or combined). Use when the user provides SMILES and asks for ECFP/MACCS/MAP4/FCFP fingerprints, RDKit or Mordred de
Designs and tunes neural network architectures, covering architecture selection, layer configuration, hyperparameters, preprocessing, transfer learning, connectivity, interpretability, blueprints, and NAS. Use when a user needs help choosing or configuring a n
Quantizes LLMs to 8-bit or 4-bit with bitsandbytes to cut memory 50-75%, covering memory estimates, loading configs, QLoRA fine-tuning, 8-bit optimizers and troubleshooting. Use when a user asks about model memory requirements, loading a model in 8-bit or 4-bi
Converts HuggingFace models to GGUF format and quantizes them with llama.cpp for efficient CPU/GPU inference. Use when the user wants to convert a model to GGUF, quantize a GGUF file, generate an importance matrix, build llama.cpp for their hardware, run llama
Guides 4-bit GPTQ quantization of large language models, including configuration, loading pre-quantized models, kernel backends, and QLoRA fine-tuning. Use when the user asks to quantize a model, load a GPTQ model, choose quantization settings or backends, or
Loads, preprocesses, analyzes, and models whole-slide pathology and multiplex imaging data with PathML. Use when the user provides a slide path, asks for stain normalization or tissue detection, wants spatial graphs, nucleus detection, model training, CODEX/Ve
Generates PennyLane code for building, training, and running quantum circuits across simulators and hardware, including variational algorithms, quantum machine learning, and molecular simulations. Use when the user asks for circuit construction, hybrid quantum
Guides GRPO/RL fine-tuning of language models with TRL for reasoning and structured tasks, covering dataset prep, reward functions, configs, training setup, and log diagnosis. Use when the user wants to prepare data for GRPO, design or test reward functions, g
Fine-tunes and aligns language models with TRL methods (SFT, DPO, PPO, GRPO, reward model training) and evaluates the results. Use when the user wants to train on instruction or preference data, run RLHF pipelines, or test an aligned model on prompts.
Use when the user asks to design a RAG pipeline, choose a chunking strategy or embedding model, pick a vector database, or evaluate retrieval quality (precision@k, recall@k, NDCG). Examples: 'design a RAG system for our docs', 'what chunk size should I use for
Manages a local Chroma vector database for creating collections, adding documents with embeddings and metadata, semantic search, retrieval, updates, and deletion. Use when the user wants to store embeddings, run semantic search or RAG retrieval, or manage Chro
Manages a Qdrant vector database for RAG and semantic search: creates collections, upserts points with payloads, runs filtered and batch searches, and retrieves collection info. Use when the user asks to set up a Qdrant collection, insert or update vectors, se
Generates local sentence-transformers embeddings for RAG, similarity, semantic search and batch encoding, and advises on model choice and vector store integration. Use when the user asks to embed texts, compare similarity, rank a corpus, batch-encode, pick a m
Hunts vector-store and embedding-layer weaknesses in RAG pipelines — persistent corpus poisoning, cross-tenant vector IDOR, source-text and metadata leakage, and retrieval hijack — with proof gates. Use when testing an authorized RAG application's retrieval la
Adds programmable safety rails to LLM applications at runtime, covering jailbreak detection, input/output validation, fact-checking, PII masking, toxicity detection, LlamaGuard integration, parallel checks and threshold tuning. Use when a user wants to block j
Computer vision engineering skill for object detection, image segmentation, and visual AI systems. Covers CNN and Vision Transformer architectures, YOLO/Faster R-CNN/DETR detection, Mask R-CNN/SAM segmentation, and production deployment with ONNX/TensorRT. Inc
