neerajcodz/ml-stack
Requirements-driven ML research, experiments, validation, and evidence audit for coding agents.
Orchestrate an end-to-end ML project when a user asks to discover a repository, lock requirements, audit data, research papers, choose a model, train, evaluate, promote, or audit a deliverable; route each phase to the owning sibling skill and stop at approval boundaries.
Independently audit ML release and reproducibility claims when a user asks whether a model, dataset, run, artifact, or deployment is ready; reconcile provenance, hashes, privacy, licenses, and gates with PASS, FAIL, or ABSTAIN.
Route ML workloads when a user asks about local CPU or GPU, HF Jobs, Sandbox, Spaces, Kaggle, Modal, SSH, Slurm, quotas, cost, logs, retries, cancellation, artifacts, or resume; plan provider-neutral execution with explicit approvals.
Inspect local or remote datasets when a user asks about schema, splits, samples, statistics, labels, leakage, privacy, licensing, SFT, DPO, GRPO, or vision annotations; produce a formatter or preprocessor handoff without inventing remote execution.
Prepare an ML artifact for serving when a user asks to deploy, publish, expose an endpoint, launch a Space or local service, configure auth, scale, smoke-test, or roll back; separate release approval from training and evaluation.
Evaluate ML models or systems when a user asks for benchmarks, metrics, slices, calibration, robustness, LLM evaluation, smoke tests, pass/fail thresholds, or go/no-go decisions; preserve revisions and uncertainty.
Plan or review an isolated ML experiment when a user asks to test a hypothesis, compare runs, tune one variable, reproduce a result, or lock validation; prevent self-promotion and preserve reproducible evidence.
Manage Hugging Face or repository artifacts when a user asks to browse revisions, inspect cards, upload files, commit, open a PR, create a discussion, configure a webhook, or publish a model, dataset, Space, or trace; keep writes approval-gated and URLs revision-pinned.
Choose an ML model or provider when a user asks which checkpoint, router, local model, adapter, tokenizer, or inference backend fits a task; compare revisions, license, context, cost, latency, quality, and training compatibility.
Perform paper-first, source-fresh ML research when a user asks for papers, methods, benchmarks, recipes, citations, or evidence-backed model guidance; corroborate claims and hand ranked recommendations to model, data, or training owners.
Track and diagnose ML runs when a user asks for run identity, metrics, artifacts, dashboards, structured events, NaN or divergence alerts, stalls, privacy, or retrieval; use ledger semantics without claiming an unavailable provider integration.
Design advanced ML training when a user asks for SFT, DPO, GRPO, reward modeling, LoRA, QLoRA, sentence-transformer, vision, checkpointing, Trackio, or HF persistence; emit an executable specification unless an equivalent runtime operation exists.