Haibarakiku

Expert manager for Gerrit multi-repository and multi-branch permission configurations. Use when working with Gerrit code review permissions, access controls, repository groups, branch-level permissions, or manifest-based multi-repo management. Use when: gerrit, permissions, code-review, access-control, devops.

CUDA expert: GPU kernel programming, memory management (global/shared/local), warp divergence, stream concurrency, cuBLAS/cuFFT integration. Use when writing GPU-accelerated code with CUDA.

Hugging Face expert: Transformers, Datasets, PEFT (LoRA/QLoRA), model fine-tuning, GGUF quantization, Text Generation Inference, pipeline optimization. Use when working with pretrained models, fine-tuning LLMs, or building NLP applications.

Jupyter expert: magic commands, nbconvert, JupyterLab extensions, remote setup, ipywidgets, profiling, debugging, cell decorators, papermill for automation. Use when working with Jupyter notebooks, data exploration, or building ML experiments.

LangChain expert: LCEL (LangChain Expression Language), chains, agents, RAG pipelines, tool calling, memory, callbacks, output parsers, retrieval strategies. Use when building LLM applications, RAG systems, or AI agents with LangChain.

Invoke when: User needs help with LlamaIndex RAG pipelines, index types, query engines, or vector stores. Provides: Index selection, embedding configuration, retrieval strategies, and pipeline optimization.

LLM serving expert: vLLM, TensorRT-LLM, Triton Inference Server, quantization (INT8/FP8/GPTQ/AWQ), continuous batching, PagedAttention, KV cache management. Use when deploying LLMs for inference.

MLflow expert: experiment tracking, model registry, autologging, MLflow Projects, MLflow Models, model serving, A/B testing, feature store integration. Use when tracking ML experiments, managing models, or deploying ML models with MLflow.

Generate an Abaqus FEA training dataset for surrogate / ML models. Latin Hypercube Sampling (or sparse-pattern sampling) over a parameterized design vector, one case folder per sample, batch-submit Abaqus jobs via subprocess, recover from crashes, and write a unified dataset index. Use when the user wants to "build a training set for a surrogate model", "sweep design parameters in Abaqus", "run N FEA simulations", or "sample a design space".

Closed-loop inverse-design validation. Given a target deformation field, solve the inverse problem on a trained surrogate (Ridge / linear), then run an Abaqus FEA verification and compare surrogate-predicted vs. true displacement field. Reports MSE / MAE / max-abs-error / NRMSE side-by-side, plus saturated-channel count, so you can quantify the surrogate-FEA gap. Use when the user wants to evaluate "is my surrogate good enough for inverse design?", "how big is the surrogate-FEA gap on this target?", "did the optimizer find a real solution or just a surrogate hallucination?"

Convert per-case Abaqus FEA outputs into ML-ready (X, Y) wide-table CSVs. Pivots irregular FEA mesh node displacements onto a regular N×N grid via direct binning (structured mesh) or bilinear resampling, picks the final frame as the deformation target, and aggregates across many cases into X_amplitude.csv (design vectors) + Y_grid_uz.csv (flattened grid displacement). Use when the user has a folder of completed FEA cases and wants to train a Ridge / MLP / Gaussian Process surrogate on the (input → displacement field) mapping.