Measuring, at every step of the path from raw data to a decision-supporting finding, how many different numbers two defensible decisions produce for the same question: the data each kind of analytics requires, the summary each measurement scale permits, the rows a data quality threshold eliminates, transformation with formulas and pivot tables in a spreadsheet and the limits of the spreadsheet, data collection, cleaning and transformation decisions, analysis with array and data frame interfaces, exploratory analysis and applied statistics, chart design, reporting and analysis at scale.
Reading the number a model produces together with the number a procedure dumber than it produces on the same measure: the labels each kind of learning asks for, the two separate numbers measured on the training set and on a held-out set, the difference a feature adds against what leakage inflates, the share of the model family and the tuning budget, what the baseline is when there are no labels, what capacity adds and what training practice carries, the baselines of the language and vision domains, what a random policy and the most popular recommendation come to, and who the baseline was chosen for.
Measuring the apparatus built around a component you did not train, cannot see inside of, and that can give two different answers to the same input: the number of the bare call, what the apparatus adds, and what the apparatus costs as tokens, calls, steps and share of failures. The surface of the model and its token cost, prompt design and reasoning techniques, building production-quality applications with model endpoints, retrieval-augmented generation with embedding vectors, the architecture and memory of tool-using agents, the structure of the tool server protocol, and measuring the defense layers against the attack surface specific to language models.
Measuring a pipeline that reports when it crashes but says nothing when it carries wrong data: a data pipeline's number is not the rows it carries, it is how many steps later a wrong row is caught. Turning the stages of the lifecycle and the quality dimensions into measurable checks, the effect of the storage layout on query cost across the warehouse, lake and lakehouse approaches, the rerunnability of ingestion, transformation and orchestration, partitioning and performance in distributed processing, the reproducibility and deployment of model training, and the governance of a model in production with drift monitoring. In every lesson the unchecked pipeline, the checked pipeline and the cost of the check itself stand side by side.
A curriculum that establishes iOS application development through the type system, value and reference semantics and memory model of the Swift language, the structure and development tools of the iOS platform, declarative interfaces with SwiftUI, UIKit and bridging, the data, networking and concurrency layer, and architecture, testing and release.
A curriculum that establishes Android application development through the null safety, class kinds and collection model of the Kotlin language, the component structure and development tools of the Android platform, declarative interfaces with Jetpack Compose, architecture components and the data layer, platform services, and testing, performance and release.
A curriculum that teaches the approaches to building for several mobile platforms from a single codebase, the limits of code sharing, bridge-based frameworks and frameworks with their own rendering engine, and the production practice of a shared codebase.