Data and Machine Learning

What the job is actually like

Overwhelmingly: finding data, working out why it is wrong, and cleaning it. Then writing queries, building something that answers a question, and explaining the result to someone who will act on it. Actual model training is a small slice, and in most jobs a well-chosen simple model beats an interesting complicated one. Communicating a finding clearly matters as much as producing it.

A roadmap

Lengths are what this typically takes alongside other commitments, not a promise. The "prove it" line matters more than the timeline — that is what someone hiring will look at.

1. SQL, to a level that surprises people

   Joins, window functions, aggregation, query plans. This is the single most transferable skill on this path and the one most often skipped in favour of flashier material.

   Typically

   One to three months

   Prove it

   Answering a real question from a public dataset in SQL alone, with the query readable by someone else.

2. Statistics you can actually apply

   Distributions, sampling, what a confidence interval does and does not mean, why correlation misleads. Enough to notice when a result is an artefact rather than a finding.

   Typically

   Two to four months

   Prove it

   Being able to explain why a plausible-looking conclusion in a public analysis is unsound.

3. Python for data work

   pandas, numpy, and plotting. The goal is fluency at reshaping and inspecting data, not memorising library APIs.

   Typically

   Two to three months

   Prove it

   A notebook that takes raw, messy public data through to a clear answer, with the cleaning decisions written down and justified.

4. Machine learning fundamentals

   Regression, classification, trees, and above all evaluation — train/test splits, overfitting, leakage, choosing a metric that matches the actual decision.

   Typically

   Three to six months

   Prove it

   A project where you explain why your model is only as good as it is, rather than one claiming suspiciously high accuracy.

5. Getting it in front of someone

   A dashboard, a scheduled job, or a small deployed service. Analysis that nobody can act on has no value, and this is the step that separates candidates.

   Typically

   One to two months

   Prove it

   Something a non-technical person can open and understand unaided.

What AI has changed

Writing a first-draft query or a plotting snippet is now close to free, which removes a chunk of what junior analysts spent their time on. What has grown in value is knowing which question is worth asking, whether the data can honestly support the claim being made, and spotting the leak or the sampling problem that makes a beautiful result meaningless. Models are also easier to build and no easier to evaluate — which means careful evaluation is now the scarce skill, not model building. Treat any generated analysis as a claim to be checked, because it will be confidently wrong in ways that look right.

Common mistakes

• Starting with deep learning. Most real jobs need SQL and clear thinking, not neural networks.
• Only ever using clean teaching datasets, then being lost when real data arrives broken.
• Reporting a high accuracy figure without checking for leakage. It is usually leakage.
• Producing analysis nobody asked for and nobody uses.
• Confusing "data scientist" with "research scientist" — the second needs a PhD and there are far fewer roles.