ai-ml-workflow

AI/ML Workflow

Overview

The disciplined lifecycle toilact follows for ML work. This skill is the map: it names the phases, the order, and which guard/doc to pull for each. Detailed playbooks live in the on-demand library — load them with /gc-lib ml-lifecycle/<phase> only when you reach that phase (they are not auto-loaded).

When to Use

Any ML/data task: EDA, feature engineering, preprocessing, training, hyperparameter tuning, evaluation, the iterate loop, on-device/service integration, deployment. Works for any ML project — tabular, time-series / forecasting, or NLP (e.g. a daily prediction pipeline, a fine-tuned transformer).

The lifecycle (follow in order; don't skip ahead)

1. EDA → /gc-lib ml-lifecycle/eda — understand distributions, missingness, leakage risk, target balance.
2. Feature engineering → /gc-lib ml-lifecycle/feature-engineering — MUST honor ml-data-leakage-guard.
3. Preprocessing → /gc-lib ml-lifecycle/preprocess — fit transforms on the train fold only.
4. Training → /gc-lib ml-lifecycle/training — reproducible runs, seeds, logged metrics.
5. Tuning → /gc-lib ml-lifecycle/tuning — search with an inner split; MUST honor walk-forward-guard.
6. Evaluate + loop → /gc-lib ml-lifecycle/loop — out-of-sample only; decide iterate vs ship.
7. Integrate → /gc-lib ml-lifecycle/integrate — share ONE feature builder train↔serve; MUST honor model-feature-versioning.
8. Deploy → /gc-lib ml-lifecycle/deploy — version, monitor, roll back safely.

Non-negotiable guards (auto-active skills)

• ml-data-leakage-guard — feature(T) uses only data known at T.
• walk-forward-guard — train-on-past / test-on-future, no shuffle.
• model-feature-versioning — frozen feature contract, no train/serve skew.

Behaviour

Follow the godchi gate: brainstorm the approach, write a short plan, confirm with the user before changing a chosen modelling decision, review results before declaring success. Reply in Vietnamese.