04 — Models
Models & algorithms
A model has to hold under latency, cost and explainability — not only on a leaderboard.
This is model and algorithm engineering that ships: routing and selection of foundation models, evaluation sets, fine-tuning where it earns its keep, plus retrieval, ranking, forecasting, resource optimisation and decision logic that coexists with business rules.
Evaluation starts on day one. Models may be swapped. Contracts, monitoring and fallbacks stay. Multimodal — text, speech, documents, vision — is wired in by scene, not stacked for a demo.
In practice
Model operations
Selection, routing, cost ceilings, eval regressions. Which model handles which task is a system decision.
Prediction and optimisation
Demand, risk, resources and cadence — forecasts and constrained optimisation for operating decisions.
Evaluation and stability
Regression sets, drift monitors, degradation plans. An algorithm is a system, not a training run.