DigitableCourses
Portal settings
Русский
Show me around

Local and account-free. No account, no sign-up: whatever you type into the tools stays in this browser's localStorage and never reaches a server.

Our own counter records page opens and finishes: the address leaves, plus one tenth-of-the-text figure. No cookies, no outside trackers, your IP is not stored, Do Not Track is honoured. How to check that

The portal's own repository is not published, so we do not call it open source. What is open:

Digit flang Agent gate tools-core IT Tools Chat digitwm blender-mcp Hugging Face Licences

The portal lives on donations, paid consultations and requested write-ups, and on Workbench sales.

Support Order a write-up Workbench

There are no plans to make the courses paid.

Career

Data and machine learning

Data pipelines, models, getting models into production and an honest assessment of their quality.

Roles and moves Self-assessment

Step 2 of 2 · Data and machine learning

Which level inside the role?

Where people go from here: Backend developer, Systems and business analyst, Product manager, SRE and platform engineering

Not sure — open the whole role card Pick another role

  1. Entry
  2. Core level
  3. Senior and beyond

Entry

Straight to the self-check

Compared with studying, both the question and the data change: the question is stated by somebody other than a teacher, and the data is always dirty.

Do now

  • I get the data myself and check where every field came fromI know which fields in the source lie, and I say so before drawing conclusions.
  • I start with a simple model and compare everything else against itI have a baseline that had to be beaten.
  • I check missing values, outliers and target leakage before trainingI can show that a feature is not peeking at the answer.
  • I write the analysis up so that it repeats on another machineThe result reproduces somewhere other than my laptop and later than today.

Where next

Core level

Sideways out of the role

What this step asks of you

  • I use probability and statistics, not only the meanI understand how a distribution differs from a mean, and why a sample can be unrepresentative.
  • I get the data myself and understand where it came fromI write non-trivial SQL and know which fields in the source lie.
  • I write reproducible analysis code, not a throwaway notebookMy result repeats on another machine and a month later.
  • I prepare features and understand that the data is always dirtyI check for missing values, outliers and target leakage before training.
  • I start with a simple model and compare everything against itI have a baseline that had to be beaten, not only a neural network.

Common mistakes

  • Treating the mean as the answer: the distribution hides both the tail and the skew of the sample.
  • Believing a metric out of a notebook: without an honest split of the sample it is inflated.
  • Dragging a neural network in where a rule was enough: the maintenance costs more and there is nothing to show for it.

What to learn

  1. Entry
  2. Core level
  3. Senior and beyond

Core level

Straight to the self-check

The model leaves the notebook: behind it come a pipeline, on-call and the quality of the data, which breaks more often than the model does.

Do now

  • I put checks on the data mart so that a break is found before the business finds itThe check fires before the question why are the numbers different.
  • I make recomputing yesterday an ordinary operationRestarting the pipeline needs no repair by hand.
  • I take the model to production and watch it after the launchI know when the model has gone stale, and I have a plan for retraining.
  • I count the cost of a request to a language model together with the quality of the answerThe solution has both an evaluation of quality and a price per answer.

Where next

Senior and beyond

Sideways out of the role

What this step asks of you

  • I evaluate a model honestly, including where it is wrongI pick the metric that fits the problem rather than the one that looks nicer, and I know the price of each kind of error.
  • I build data pipelines that survive a source going downRecomputing yesterday is an ordinary operation, not a repair by hand.
  • I answer for the quality of the data, not only for the code that reads itThe data mart has checks, and they fire before the business notices.
  • I take a model to production and watch it after the launchI know when a model has gone stale, and I have a plan for retraining.
  • I work with language models as an engineer, not by guessing at promptsMy solution has an evaluation of quality and an estimate of the cost per request.

Common mistakes

  • Improving the metric instead of the problem: the number goes up and the benefit stays where it was.
  • Forgetting about the data after the launch: the source changes silently and the model quietly rots.
  • Building a pipeline with no way to recompute: any failure in the source turns into work by hand.

What to learn

  1. Entry
  2. Core level
  3. Senior and beyond

Senior and beyond

Straight to the self-check

The question you answer changes: not which model is better but whether a model is needed here at all — and no is also an answer.

Do now

  • I check whether a model is needed before training oneThere are problems where I proposed a plain rule instead, and it was accepted.
  • I write the rules of an experiment down before it starts: what we count and when we stopThe result is believed because the conditions were known in advance.
  • I name the harm a mistake by the model does and what we have limited it withI can say who a mistake hurts and what happens next.
  • I tell the result to people who do not read chartsAfter the talk a decision gets made rather than a request to send the slides.

Where next

Not another step — a fork

Sideways out of the role

What this step asks of you

  • I translate a business problem into a learning problem — or say honestly that no model is needed hereI have projects I proposed to close because a rule with three conditions was enough.
  • I set up an experiment so that its result can be trustedI compute the power in advance and know what peeking at interim results does.
  • I keep the load and the cost of model inference under controlI know what a single request costs and can bring that number down.
  • I see the risks of a model: the data, privacy, the harm from an errorI can name who my model's error harms and how, and what we did about it.
  • I explain a result to people who do not read chartsAfter my talk a decision gets made, rather than a request to send the slides over.

Common mistakes

  • Peeking at interim results and stopping the experiment at a convenient point: the conclusions fall apart.
  • Hiding uncertainty: a single number with no spread reads as a promise.
  • Arguing about model architectures instead of the problem: the customer needs an answer, not a comparison.

What to learn

IT Tools source · GPL-3.0 Digitable Chat source · GPL-2.0-or-later Legal documents