Data Strategy and AI-Ready Infrastructure
From data assets to AI-ready data
AI-ready data is fit for a specific purpose: it has known lineage, defined ownership, documented quality, appropriate access controls, and the right shape for the task (labeled for supervised learning, retrievable and current for retrieval-augmented generation, well-instrumented for evaluation). Candidates should understand the data supply chain - source systems, ingestion, transformation, feature or embedding stores, and serving - and where quality debt accumulates.
Key concepts: data contracts between producers and consumers, master and reference data, metadata and cataloging, lineage, data quality dimensions (completeness, accuracy, timeliness, consistency, uniqueness), and the difference between operational and analytical data needs. For generative systems, add document freshness, chunking and retrieval quality, and permission-aware retrieval so the model does not surface content a user cannot see.
Infrastructure topics: compute choice (training vs. inference economics), latency and throughput requirements, environment separation, MLOps/LLMOps pipelines, versioning of data and models, and cost observability. Know why inference cost and vendor lock-in are strategic, not just technical, concerns.
Traps in questions
- Treating a data lake or warehouse as sufficient; readiness is about fitness for the use case, not storage.
- Assuming more data always helps; label quality and representativeness usually matter more.
- Overlooking access control in retrieval systems, creating a leakage path.
- Ignoring data drift and upstream schema changes that silently degrade models.
Rule of thumb
No data contract, no production. Before a model goes live, the producing system must commit to schema, freshness, and quality thresholds, with an owner who is alerted on breach. Worked example: a support-assistant RAG system requires a contract on the knowledge base - nightly refresh, mandatory metadata tags, and permission inheritance - plus an evaluation set of real questions to detect retrieval regressions.