title: “AI Lab · Machine Learning Foundations” description: “How I approach classical ML inside the AI Lab stack.”
Machine Learning Foundations
Even in an LLM world, classical ML powers plenty of AI Lab workflows. Here’s the playbook.1. Problem Framing
- Define Decision Surface: what action/outcome are we predicting?
- Capture Data Reality: source, volume, freshness, labeling cost.
- Identify Success Metric: accuracy, F1, RMSE, cost savings.
2. Dataset Ops
- Ingest: Airbyte → BigQuery, version via LakeFS.
- Label: Snorkel + human-in-the-loop in Label Studio.
- Split: Temporal splits for time series, stratified for classification.
- Data cards: Document biases, schema, owners.
3. Modeling Stack
- Tabular: CatBoost, XGBoost with Optuna tuning.
- Time-series: Prophet + Nixtla NeuralForecast.
- Recommendation: implicit matrix factorization + rerank via embeddings.
4. Deployment
- Package models with BentoML, deploy to AWS Lambda / ECS.
- Monitor with Evidently AI; send drifts to Slack.
- Build guardrails: auto-disable if metrics degrade beyond threshold.
5. Use Cases
- Productivity habit scoring.
- AviWealth cashflow anomaly detection.
- Thinki.sh challenge recommendations.