Data That Can Be Trusted
Most organisations have governed data. Far fewer have data that other teams, and now AI systems, can safely build on without asking around first. This six-part series explains what closes that gap: treating important datasets as products with owners, contracts, checks and service levels; settling what business terms mean before an assistant answers from them; choosing between querying, retrieval, fine-tuning and knowledge graphs by the question being asked; building real data engineering capability; and operating data quality like any production service. One fictional Free State grain co-operative carries the examples, and all figures are illustrative.
6 articles
- Part 1A data product is more than a governed tableMost organisations already have governed tables. What they rarely have is data that other teams, and now AI systems, can safely build on without asking around first. Part 1 explains the difference: an owner, a contract, tests, a service level and known consumers.9 min read
- Part 2AI projects expose weak data ownershipWhen an AI assistant gives two customers two different answers to the same question, the cause is rarely the model. It is usually a word that four teams define four ways and nobody has the authority to settle. Part 2 explains what data ownership really involves and how to assign it.7 min read
- Part 3RAG, fine-tuning and knowledge graphs solve different problemsTeams often choose between retrieval, fine-tuning and knowledge graphs as if they were competing products. They are not. Each answers a different kind of question, and a fourth option, simply querying the data, is the right answer more often than any of them. Part 3 gives a decision method.6 min read
- Part 4Knowledge-graph search for relationship-heavy questionsSome questions are not about any single document or record. They are about how things connect: which members, which bins, which shipments. Knowledge graphs answer those questions well, at a real cost. Part 4 explains how graph search works, where it fails and when it is worth building.6 min read
- Part 5Ten data engineering projects that prove real capabilityMost data engineering portfolios show that someone completed tutorials. Very few show that they can run a pipeline that fails, recovers and stays trusted. Part 5 sets out ten projects, in order, with the business outcome and the evidence that makes each one credible, built around South African cost and connectivity constraints.6 min read
- Part 6Data quality and observability are operational disciplinesData quality is usually treated as a clean-up project. It works better as operations: checks placed where they can stop a bad load, alerts that reach the owner, lineage that tells consumers before they notice, and a review after every incident. Part 6 closes the series.6 min read
Working through any of this in your own organisation?
These articles come out of delivery work rather than theory. If the problems in this series look familiar, a short scoping conversation is usually the quickest way to work out what is worth doing first.
Talk to CloudNala