Week 6 — From Learning Models to Building Tools

The first five weeks of this AI learning journey focused on understanding how models behave.

I explored classification, imbalanced datasets, accuracy, precision, recall, ROC curves, Precision-Recall curves, and thresholds.

Those experiments gave me useful foundations.

They also reinforced something I keep coming back to: the interesting part is rarely the model in isolation. It is how the model, the surrounding logic, and the human workflow come together.

So for the next phase of the journey, I want to shift slightly.

I want to start building things.

From experiments to small tools

I do not want the AI Learning Lab to become a collection of disconnected notebooks.

The experiments matter because they help build intuition.

But ultimately, I want that learning to lead to practical systems.

The next stage will therefore focus on small, focused tools that solve specific problems.

Not large platforms.

Not vague AI demos.

Small tools with clear inputs, useful outputs, and a real workflow around them.

Starting with a problem I already understand

The first area I want to explore is job search.

Over the past few months, I have repeatedly worked through tasks such as:

• understanding what a role is really asking for
• comparing a job description with my experience
• identifying strong matches and genuine gaps
• deciding which parts of a CV to emphasise
• tailoring experience without misrepresenting it
• preparing cover letters
• identifying likely interview themes

These tasks are repetitive enough to benefit from tooling, but nuanced enough that human judgement still matters.

That makes them interesting candidates for applied AI.

The first tool: Role Fit Analyzer

The first tool I want to build is deliberately simple.

Given a job description and a CV or experience summary, a Role Fit Analyzer should help answer questions such as:

• What capabilities is the employer actually looking for?
• Where does the candidate's experience align strongly?
• Where are the gaps?
• Which experience should be emphasised?
• What evidence supports each claimed match?
• What areas are likely to come up during interview?
• What should the candidate investigate further before applying?

The objective is not to produce an arbitrary compatibility score and declare whether someone should apply.

The objective is to make the analysis more structured and explainable.

Why not just ask an LLM?

That is actually part of what I want to learn.

It is easy to paste a CV and job description into an LLM and ask:

“Am I a good fit?”

But building a useful tool means thinking more carefully about the workflow.

For example:

• What information should we extract from the job description?
• How should requirements be classified?
• How do we distinguish direct evidence from inferred alignment?
• How should missing experience be represented?
• How do we prevent the system from inventing qualifications?
• What should be deterministic, and what should be generated by a language model?
• How do we make the output useful enough for someone to act on?

Those questions turn a prompt into a system.

And that is the part I am interested in exploring.

Connecting this back to the first five weeks

This is not a departure from the earlier learning.

It is the next layer.

Weeks 1 to 5 were largely about understanding model behaviour and evaluation.

That taught me to think about:

• inputs
• outputs
• thresholds
• errors
• trade-offs
• explainability
• the consequences of decisions

Those same principles matter when building AI-assisted products.

A useful tool needs more than a model or an API call.

It needs a clear purpose, sensible boundaries, understandable outputs, and a human who remains in control of the final decision.

Building the first version

I want the first version to stay deliberately small.

The initial workflow will probably look something like this:

1. provide a job description
2. provide a CV or structured experience profile
3. extract the key requirements from the role
4. map evidence from the candidate's experience
5. identify strong matches and gaps
6. produce a structured analysis
7. suggest areas for further preparation

That is enough for a useful first build.

If it works, later iterations could explore:

• CV bullet tailoring
• cover letter preparation
• interview question generation
• application tracking
• comparison between multiple roles

But those are later problems.

For now, the goal is one small tool that does one job well.

This Week’s Learning

The first part of this journey was about learning through experiments.

The next part is about learning through building.

The principle I want to keep is the same:

Start small enough to understand what is happening.

For the next few weeks, that means turning practical problems into small tools, testing whether they are actually useful, and documenting what I learn along the way.