Architecting Intelligence
A build-led learning lab exploring how useful AI-assisted software is designed, evaluated, and improved.
I've spent years working inside regulated financial ecosystems — payments infrastructure, scheme alignment, partner integrations, and delivery under strict governance. In those environments, how a system behaves matters as much as what it can do.
I'm applying that same discipline to machine learning and generative systems. This isn't a pivot away from systems engineering — it's an extension of it.
Architecting Intelligence started with experiments designed to understand model behaviour. It is now evolving into a build-led learning lab: small, useful tools built around real workflows, with each build creating new questions about architecture, evidence, evaluation, and human decision-making.
This page documents that process — what I build, the decisions behind it, what I test, what fails, and what I learn along the way.
Experiments, notebooks, and code from the learning journey live here: AI Learning Lab →
Why this page exists
Building useful AI software involves more than sending a prompt to a model. Real systems need structured inputs, clear boundaries, reliable outputs, evaluation, and ways to handle uncertainty.
They also require decisions about where deterministic software should end and model-generated behaviour should begin — and how a human remains in control when the output influences a real decision.
The purpose of this lab is to explore those questions by building rather than studying them only in isolation.
Each tool starts with a real problem. The implementation becomes the experiment, and the results determine what I explore next.
From Learning to Building
Foundation — Weeks 1–5
Understanding Model Behaviour
The first five weeks focused on understanding how machine-learning models behave rather than simply how to train them. Starting with a basic classifier, I explored imbalanced data, precision and recall, ranking, and decision thresholds.
The key lesson was that models produce scores and probabilities; the decisions, trade-offs, and consequences emerge from the systems we build around them.
Focus: classification · imbalanced data · precision and recall · ranking · thresholds · decision quality
Build Lab — Week 6 onwards
Building Useful Tools
The next stage shifts the learning into small, useful AI-assisted software built around real workflows. Instead of exploring concepts in isolation, each build becomes the environment for learning.
That means thinking beyond the model itself: how information is represented, what should be deterministic, where generated behaviour adds value, how claims are grounded in evidence, and how the resulting system should be evaluated.
Focus: structured data · evidence · deterministic vs generated behaviour · evaluation · explainability · human-in-the-loop decisions
Current Build — Role Fit Analyzer
A small AI-assisted tool for comparing a candidate's experience with the requirements of a job description.
The goal is not to produce an arbitrary compatibility score. The tool should identify relevant evidence, distinguish direct matches from reasonable inference, expose genuine gaps, and avoid inventing experience that is not present in the source material.
Building it provides a practical environment for exploring requirement extraction, structured representations, evidence matching, deterministic and LLM-generated behaviour, hallucination controls, and evaluation.
Read the Week 6 introduction →Latest Notes
Short, cumulative notes documenting what I learn, build, test, and change along the way.
Week 4 — When the Model Doesn’t Decide
2026-03-30
Week 2 — Looking Beyond Accuracy
2026-03-19
Evaluation Metrics in Regulated Systems
Why accuracy is insufficient in regulated and imbalanced systems.
2026-03-09