Not a Demo. A Real System.
I used AI to design, build, deploy, and generate measurable results in a production system in under 30 minutes.
- Not a prototype.
- Not a demo.
- A real platform. With real users.
And importantly — under real constraints.
This isn’t a story about AI generating content.
It’s about what happens when you apply AI to a system that already exists, already works, and already has users — but lacks visibility.
The Context
For over six years, I’ve been building and operating a platform called WhereWeLearn.
It’s a global, charity-led initiative designed to help people discover and organise free educational content.
The platform is built on the LEAST engine (Linking Educational And Social Technologies), a production system that has evolved over 8+ years without a framework, using:
- server-rendered PHP
- domain-based libraries
- global state
- a centralised database abstraction layer
This wasn’t accidental.
The system is designed to be:
- stable
- understandable
- controllable
And importantly — usable in the real world.
The Constraint (This Changes Everything)
WhereWeLearn operates under a strict constraint:
It does not promote content.
No marketing.
>No optimisation.
>No algorithmic bias.
This is by design.
As a charity-led platform, it must remain:
- neutral
- non-commercial
- unbiased in how content is surfaced
Despite that, over time the platform generated:
- 134,000+ engagement events
- users across 130+ countries
- entirely organic discovery
Which creates an unusual situation:
👉 A system with real usage — but no structured way to measure it.
The System Behaviour
Delivering on the strategic learning goals, firstly each lesson created becomes:
- an indexable asset
- part of a structured learning graph
- connected to materials and related lessons
Specifically the platform relies on:
- sitemap accuracy
- OpenGraph integration
- internal linking
- distributed entry points
Users don’t arrive through a homepage.
They arrive:
- via search engines
- directly into materials
- inside specific learning contexts
In effect, this is programmatic SEO without marketing.
The Problem
The system worked well.
But I couldn’t answer basic questions:
- Are users engaging deeply or bouncing?
- Are they following learning paths?
- Is the system improving over time?
There was no meaningful feedback loop.
Which means:
The system could evolve — but not intelligently.
The Intervention
Instead of rebuilding anything, I introduced two things:
1. AI-Assisted Engineering
Using tools such as Anthropic Claude Code and ChatGPT, I:
- designed a measurement model
- implemented a reporting engine
- deployed it into production
Time from idea → live system:
Under 30 minutes
2. A Measurement Layer (Without Tracking Users)
Rather than introducing cookies or heavy analytics, I extended the existing audit system through AI and measurable requirements.
The principle:
Track behaviour, not people.
This enabled:
- session depth approximation
- lesson vs material engagement
- learning flow tracking
- bot filtering
- time-based measurement
All within the existing architecture.
The First Output
Furthermore within minutes of deployment, the system produced measurable data:
- Average session depth: 7 pages
- Lesson engagement: 26.1%
- Learning flow rate: 8.3%
- AI-driven traffic: 0% (baseline confirmed)
Interpreting This Properly
This is early data. Consequently it needs to be treated carefully.
For example:
- The session depth was generated during system testing (not yet real user behaviour)
- Traffic volume is too small to draw conclusions
- Learning flow requires more time to stabilise
But one signal is already meaningful:
Lesson Engagement Is Increasing
Historically:
- ~3% (2024)
- ~17% (2026 YTD)
Initial measured value:
- 26.1%
Hence even allowing for small samples, the direction is clear:
Users are moving from passive consumption to structured learning.
What Actually Changed
This is the key point.
The platform itself didn’t change.
No redesign.
>No feature expansion.
>No new content strategy.
What changed was:
1. The System Became Observable
Before:
- behaviour unknown
After:
- behaviour measurable
2. The Feedback Loop Collapsed
Before:
- weeks to understand impact
After:
- minutes
3. AI Became an Activation Layer
Not replacing engineering.
Not replacing systems.
But enabling:
- faster iteration
- better decisions
- measurable outcomes
What Didn’t Change
This matters just as much.
- The platform remains neutral
- No promotion occurs within WhereWeLearn
- No user tracking was introduced
- Governance constraints remain intact
Activation happens externally.
Measurement happens internally.
The system remains trusted.
What This Means
Furthermore this approach isn’t specific to this platform.
It demonstrates something broader:
AI doesn’t need to replace systems to be valuable.
It needs to make them understandable.
Equally once a system is measurable:
- decisions improve
- iteration accelerates
- outcomes become visible
What Comes Next
This is still the baseline.
The next step is controlled activation:
- introducing AI-driven content externally
- directing traffic through philipalacey.com
- measuring the impact on behaviour
And most importantly:
Observing what actually changes — and what doesn’t.
Final Thought
Most discussions about AI focus on capability.
In practice, what matters is:
- where it is applied
- what constraints exist
- whether impact can be measured
In this case:
AI didn’t replace the system.
It made the system measurable.
And that’s where real change begins.
