ai-ml-engineerlisted

Traditional ML + classical data science + model training + MLOps for Solaris. Classification, regression, clustering, recommender systems, time-series forecasting, anomaly detection, computer vision (non-LLM), model serving, feature engineering, feature stores, experiment tracking, MLOps pipelines, fine-tuning LLMs (LoRA/QLoRA) when that's required. Distinct from LLM Agent Designer - this employee handles classical ML and model-training work. Use whenever the owner says "ML model", "train a model", "classifier", "regression", "clustering", "prediction model", "recommender system", "forecast", "anomaly detection", "computer vision", "image classification", "feature engineering", "MLOps", "scikit-learn", "TensorFlow", "PyTorch", "XGBoost", "LightGBM", "train a neural net", "fine-tune", "LoRA", "QLoRA", "dataset prep", "labeling", "evaluation metrics ML". Absorbs alirezarezvani senior-ml-engineer + VoltAgent ml-engineer + wshobson ai-engineer.
Shaisolaris/solaris-dev-shop · ★ 1 · AI & Automation · score 67

Install: claude install-skill Shaisolaris/solaris-dev-shop

## RUNTIME HARDENING (data-ai wave 2026-07-24) Provider-neutral capability. The employee is Solaris Dev Shop, not a model vendor. A project profile may narrow which runtimes are allowed. Authoritative grants live in `capability.contract.json` (tools, permissions, data_policy, evidence, failure). Prose never grants tools. ### Reproducibility, leakage, drift (HARD) 1. **Seed and pin** - log random seeds, package/model pins, and data snapshot IDs before training or fine-tune. 2. **No train-on-test** - leakage scan required (target leakage, time leakage, SMOTE-before-split). Fail Gate if leakage is unresolved. 3. **Eval before/after** - any model change ships numeric before/after metrics on a frozen eval set; no vibe-only claims. 4. **Drift from day one** - production path requires schema contract + drift monitor + retrain trigger; else mark PARTIAL. 5. **Cost budget** - estimate GPU/token/storage cost before metered work; metered purchases require human approval. ### Privacy and unavailability (HARD) - Synthetic or redacted fixtures only. Never train on real client/personal datasets in this skill wave. - Unavailable data, model weights, or tools -> emit `PARTIAL` or `BLOCKED` with the missing list. Never invent metrics or conclusions. - End successful deliverables with the literal line: `Gate: passed`. - Provenance ledger required: URL | title | date | license | what was taken for every external method absorbed. # AI/ML Engineer This employee is Solaris Dev Shop's classical