Skip to content

Paul Wilson

AI Engineer · AI-powered websites and apps

Recruiting for a role? Check me against your job ad. Three AI agents read it and give you an honest match, gaps included.

Last updated October 2026

Profile

I build AI-powered websites and applications, taking full ownership from the initial idea through to a finished, live product. I'm methodical and deliberate in how I work: I start with a clear understanding of the desired outcome, break projects into manageable sections, and test each part before moving on. This allows me to iterate quickly while maintaining a high standard of reliability, usability and visual detail.

I've been building with LLMs for around three years, most recently creating Vantage City OS, a live 3D map of London built using Ordnance Survey and open data. I'm curious about new ideas and particularly enjoy finding practical ways AI can make a meaningful difference to a product or workflow. I'm also comfortable explaining technical concepts simply and working from an initial idea through to a working prototype.

I use modern AI development tools including Claude Code to rapidly build, test and refine websites, applications and prototypes. I have a strong eye for detail and take pride in producing work that not only functions well, but feels polished and considered. I'm comfortable integrating with unfamiliar systems and picking up new technologies quickly.

Behind this is 25 years of experience in enterprise IT, giving me a broad technical foundation and the ability to understand both the bigger picture and the details that make a product work.

Core skills

AI & LLMs
Generative AI, large language models (LLMs), prompt engineering, structured outputs, AI agents, multi-agent orchestration and tool calling, MCP (Model Context Protocol), RAG over unstructured documents (embeddings, chunking, vector search with Pinecone), prompt caching, streaming, and reliability (model fallbacks, retries, time limits). Anthropic Claude API, OpenAI, Gemini and Groq, Hugging Face models through the inference API or run locally in the browser as ONNX.
Front End & 3D
TypeScript, JavaScript (ES6+), React, Next.js (App Router), Three.js and React Three Fiber (WebGL), Tailwind CSS, HTML/CSS, responsive design, UI and interaction design, design systems, no-code and low-code (Bubble).
Creative & Design
Design prototyping, creative coding, interactive 3D, Figma, AI image, video and voice tools.
How I Build
Vibe coding and AI-assisted development with Claude Code, Cursor and MCP servers, working from a written spec. Agentic coding, rapid prototyping, MVPs, proofs of concept and automated testing.
Full-Stack & SaaS
Full-stack web development: Node.js, REST APIs, Supabase (Postgres, Auth, Storage), Stripe (subscriptions, webhooks), Brevo (transactional email over its REST API, deliverability), Vercel (serverless, Cron), OAuth, SEO, AEO and GEO (Google Search Console, structured data/JSON-LD, sitemaps, llms.txt, IndexNow), Git and GitHub.
Speech & Real-Time
Speech-to-text and text-to-speech (Deepgram, Cartesia, ElevenLabs, Inworld), voice cloning, WebRTC and WebSocket streaming, LiveKit, real-time face tracking (MediaPipe).
Cloud & DevOps
Google Cloud Platform (GCP) and the Cloud Console: projects, OAuth clients and consent screens, API keys and enabling services. Azure DevOps, CI/CD pipelines, PowerShell, MSBuild, InstallShield, SQL/T-SQL, IIS and Docker.

Selected projects

Vantage City OS

vantage-city.dev

A live 3D map of London drawn as holograms, built on Ordnance Survey and open data

London in 3D with live data where it happens: aircraft, Tube status, road disruptions, cycle docks, weather and air quality, and routes scored against air pollution. Ordnance Survey data adds place search, borough and ward boundaries and grid references, with a soundscape synthesised live.

Custom camera detectors, trained in the browser from a few example photos

Describe what the camera should notice, add some example photos, and a small model trains on the local GPU with TensorFlow.js. Claude flags photos in the wrong group, up to three cameras vote on what they see, and a detector can speak, notify, message Slack, Discord, a webhook or email, or be shared as a link.

Glass Box

glassbox.run

A flight recorder for AI agents, with real recorded runs you can play back and branch

Records AI agent runs and plays them back in the browser, with a timeline, a 3D graph of the agents and the actual request and response for every tool call. At each approval point you can pick a different answer and watch the real run that followed, because every alternative was recorded too.

Loose Brief

loosebrief.dev

AI design tool that turns a short brief into brand directions, a design system and a downloadable website

Answer a few questions about a business and it returns three brand directions, each with its own colours, fonts, tone of voice and homepage. Exports a static website, a Tailwind theme and brand guidelines.

Rival Radar

rivalradar.me

AI-driven competitor tracking SaaS built with Next.js, Supabase, Stripe and Vercel Cron

A SaaS app that finds a business's competitors and watches their pricing, features and messaging, with Stripe subscriptions, scheduled checks on Vercel Cron, and Claude summarising changes only from pages it has actually fetched.

Interactive Avatar

interactiveavatar.co.uk

Live conversational avatar platform with multi-provider AI agents and real-time voice and video

A live 3D avatar, lip-synced to its speech, with a conversational agent on OpenAI, Gemini or Groq: tool calling, fallback between providers, real-time speech over WebRTC and WebSocket, voice cloning, face tracking and a developer API.

Shard Islands

shardislands.me

Real-time multiplayer 3D flight game in the browser, with a world generated entirely in code

A multiplayer flight game in the browser with the whole world generated in code. React Three Fiber on the front end, a Colyseus game server on Fly.io keeping players in sync, and automated test players checking the multiplayer timing.

Chart Chic

chartchic.uk

E-commerce site for a small business selling digital cross-stitch patterns

Next.js and Supabase, with Stripe Checkout, downloads unlocked once payment goes through, an admin area, and reviews limited to people who bought the pattern. SEO set up throughout, including structured data and generated social images.

Experience

Independent AI Engineer & Web Developer

March 2020 – Present

Independent, Bubble.io Plugin Business & Client Work (Remote)

Building AI products and websites, alongside a plugin business for the Bubble platform. Ran alongside my full-time job at Capita until July 2026, and full time since.

  • Built and published 47 commercial JavaScript plugins for the Bubble platform, released to paying developers who then report the bugs: AI agents with tool calling, RAG over documents backed by Pinecone or an ordinary database, real-time voice over WebRTC and LiveKit, and Hugging Face models run either through their inference API or locally in the browser as ONNX models in a worker, across more than 25 model tasks. Also video players, 2D and 3D mapping, native mobile components, rich text editing, service workers and Apple's Liquid Glass effect. The most popular has over 1,100 installs, and they still earn money every month. See what the AI plugins do
  • Shipped eight live AI products on my own, including Vantage City OS, Glass Box and Rival Radar. The concept, the architecture, the prompts, the testing and the deployment are mine on every one. I direct the build with Claude Code and can account for every line that ships.
  • Multi-agent features where the orchestration is code rather than a model: agents run in parallel against a shared prompt cache, every call has its own deadline inside a fixed overall budget, models fall back server-side when one is unavailable, and runs stream over NDJSON with heartbeats and cancel themselves when the visitor leaves. Nothing is reported as a match unless the quoted evidence is found word for word in the source, checked in code and not left to the model.
  • RAG over unstructured documents, built by hand rather than taken from a framework: chunking with adjustable size and overlap, OpenAI embeddings, similarity search with score thresholds and result limits, stored in Pinecone or an ordinary database, and each customer's documents kept separate so one tenant's files can never appear in another's answers. Also embedding existing relational data so ordinary app records can be searched semantically.
  • Real-time voice and vision: speech-to-text and text-to-speech through Deepgram, Cartesia, ElevenLabs and OpenAI, voice cloning, 3D avatars lip-synced from ARKit and Oculus visemes with Ready Player Me and Avaturn models, webcam face tracking over WebRTC, and small vision models trained in the browser on the local GPU with TensorFlow.js.
  • Front-end contract work for international clients including Montra, Krokit and other start-ups: application interfaces, custom mapping, video, API integrations and interactive web experiences. Most of it came from people who found the plugins.

DevOps Software Packager / System Administrator

September 2012 – July 2026

Capita (formerly Clinical Solutions)

Moved into the development team after Capita bought Clinical Solutions, to own packaging and build automation.

  • Built JavaScript web apps for internal staff to speed up everyday processes, plus watchers that kept an eye on other services the business relied on. The developers were C# specialists, so I'd often help out whenever JavaScript or MVC web app work came up.
  • Ran the company's Azure DevOps setup, led the move from on-premise TFS, and built the CI/CD build and release pipelines every product team relied on, linked to ServiceNow and Aha! through REST APIs and webhooks.
  • Automated builds, deployments and software packaging with PowerShell, JavaScript, MSBuild and InstallShield.

2nd-Line Support Analyst → 3rd-Line Configuration Consultant

December 2002 – September 2012

Clinical Solutions

Supported 3,000+ clients using clinical call-logging software, including NHS Direct and NHS 24. Promoted to 3rd-line: client configuration, installs, QA testing and T-SQL database fixes on a bespoke .NET application.

    Earlier career

    Service support analyst at Surrey County Council Trading Standards; before that, computer room operator at Watson Wyatt, bar work and a post room.

    Education & development

    Self-taught. My formal qualifications are GCSE-level. I started writing production code in Capita's development team, and everything since, from 47 published plugins to eight live products, I've learned by building real things. I keep it up to date.

    Additional

    • Keeps up with AI through ByteSize, Tech Brew, The Batch and The Rundown AI. Interests: technology and cars.