What I build, in my own words.

Context

The path from feature idea to deployment was the bottleneck, and the proposed fix was to put AI tools in front of more people.

Before

People without a programming background could get working output from a model but could not see the architecture they were changing, so they could not tell when a suggestion would break something several layers down.

Outcome

delivered Measured over an early pilot: rework by senior developers dropped from 80% to 30% of pull requests submitted by non-developers — 20 pull requests from three non-developers over three months. No production incidents from those changes over the same three months. Setup for a new person is immediate where it previously took six hours plus weeks of follow-up support. The constraint surface is still being tuned, so these are figures to date rather than a final result.

A prompt written against guidelines I supply is interpreted through the Claude SDK, planned before anything is written, and then carried out by specialised sub-agents — one for database changes, one for the API layer, one for the interface, with a coordinator sequencing them. The whole of it runs inside an explicit constraint surface: what the model may touch, what it may not, which patterns are allowed and which are refused.

The architecture was the easy half. The tuning is what took the time, and it ran on two tracks at once: the rules themselves needed cycle after cycle before they behaved, because a first version is always either too permissive or so strict it blocks legitimate work; and the people using it needed the same iteration, to find where a guardrail felt protective and where it felt arbitrary.

Demonstrates the pattern behind any attempt to put AI building tools inside an organisation — the AI Enablement engagement. The model is capable from the first day. What takes the months is the constraint surface around it, and that is where the work is.

A platform for delivering structured digital training — courses, modules, assessments, completion tracking. Built to host both human-authored and AI-augmented content, with assessment logic that adapts to learner performance rather than presenting fixed question sets in sequence.

The design principle here is that AI should be invisible to the learner — the experience should feel like a well-designed course, not like a chatbot. The AI logic lives in the assessment engine, the gap identification, and the content sequencing, not in the interface.

Demonstrates the infrastructure pattern for any "deliver structured learning" use case: internal team training, customer education, certification programs. The underlying platform is the same; only the content and assessment rules change for each use case.

A landing page from one prompt produces HTML, and the HTML is mediocre: generic headlines, weak structure, copy that reads like every other page written the same way. Longer prompts helped marginally. More context, more examples and more constraints kept returning less each time, and the output stayed generic in a way I could not engineer out of a single instruction.

Splitting the work into three prompts, each with one responsibility, changed the output visibly. A planner takes the brief and decides strategy — page type, sections, order, the dominant call to action — and returns a structured plan rather than copy. A writer takes that plan and writes only the copy. A renderer turns the copy into semantic HTML with inline SVG where a visual belongs. Each stage is narrow enough that the model does one thing well instead of balancing three.

The same shape holds beyond landing pages. Articles and long-form documents behave identically: a single instruction to produce the whole thing yields acceptable and bland, while a planning pass, a writing pass and a revision pass against a defined voice produce something worth publishing. Output quality tracks the separation of responsibilities in the prompt rather than the size of it, which is the single-responsibility principle arriving somewhere new.

A production language-learning application using LLMs for adaptive lessons, conversational practice, pronunciation feedback, and personalized curriculum. Not a wrapper around a chat API — a structured learning system where the AI adapts to learner performance across sessions and surfaces gaps in a way that a fixed curriculum cannot.

The system handles auth, billing, accessibility, and performance at consumer scale. I built this because it forced me to handle what proprietary enterprise tools often skip: real user sessions at unpredictable rates, aggressive latency requirements, mobile-first layout decisions, and graceful degradation when model APIs are slow.

Demonstrates the ability to ship a real consumer-facing AI product end-to-end — including the parts no one puts in a demo: error handling, fallback behavior, cost optimization under real usage, and the operational posture required to keep it running reliably.

A platform that generates structured digital products — illustrated books, sticker packs, posters, activity books — from high-level inputs, with prompt patterns embedded into the workflow rather than exposed to the operator. The system handles format assembly, style consistency, and output packaging automatically.

The design challenge here is that "generate creative content" is vague in a way that makes most AI pipelines inconsistent. The interesting work is in constraining the generation space so that output is consistent and on-brand without requiring the operator to be a prompt engineer.

The creative production pattern applies far beyond the consumer use case. The same approach — templated generation with embedded domain knowledge — works for internal marketing assets, sales collateral, training materials, and branded templates at any volume.