From a single AI API to a full ecosystem of working products — built through vibecoding, rapid iteration, and real-world testing.
The Lumination AI API is a versatile inference platform designed for education, training, and knowledge management.
As my team developde the API on a very robust, secure and compliant backend, I used it as the foundation for a series of vibecoding experiments, building functional AI tools that solve real problems, from interactive homework helpers to corporate training simulators. Each experiment pushed the API in a different direction, and every one shipped as a working product.
This page documents the journey: the API itself, the experiments built on top of it, and the WordPress plugins now running in production.
The Foundation: Lumination AI API
Lumination provides a unified API for multiple AI capabilities under a single authentication layer. Rather than juggling multiple AI providers, a single API key unlocks:
- Conversational AI — a stateless chat agent that accepts system instructions, message history, and page context to power chatbots and assistants
- Upload & Solve — image and document understanding that extracts text from photos, PDFs, and handwritten notes, then reasons through the content
- Follow-up Tutoring — contextual conversation threads that build on previous answers for deeper learning
The API returns structured responses including token counts and credit usage, making it straightforward to build analytics, rate limiting, and cost tracking into every product.
What makes this interesting from a product perspective is the versatility: the same API powers a floating chatbot on a café website, a step-by-step maths solver for students, and a leadership training simulator — three very different user experiences, all from the same foundation.
The Experiments: ai.azvai.com
The Azvai Experiments site is a living laboratory — a collection of interactive AI tools built with Astro, vanilla JavaScript, and the Lumination API. Each experiment is a self-contained product that explores a different use case.
Leadership Decision Simulator
A multi-step interactive training tool where users practice real-world management decisions.
The flow works like this: users configure their role, challenge type, and difficulty level. The API generates a realistic workplace scenario with multiple response options. After the user makes their decision, the AI provides a coaching debrief with scoring and analysis of their reasoning.
This experiment tested the API’s ability to generate structured, scenario-based content — not just free-form text, but formatted decision trees with consistent output that the frontend could parse and display as an interactive wizard.
AI Coach Chatbot
A lightweight ChatGPT-style interface with a twist: users can upload PDFs, images, and written notes as context before starting a conversation.
The two-step wizard — first upload context, then chat — demonstrates how the API handles multimodal inputs. The coach can reference uploaded materials while answering questions, making it useful for studying, document review, or preparation for meetings.
PDF Quiz Generator
Upload a document, configure the number of questions and difficulty level, and the API generates a multiple-choice quiz with explanations for each answer.
This pushed the API’s structured output capabilities. The response needs to be a valid quiz — questions, options, correct answers, and explanations — all consistently formatted so the frontend can render an interactive assessment experience.
Technical Architecture
The experiments site is built with Astro 5 (static-first with server-side rendering), deployed on Netlify. All API calls happen server-side through Astro API routes, keeping the API key secure. The frontend is entirely vanilla JavaScript — no React, no Vue — keeping the bundle minimal and the interactivity fast.
Key patterns that emerged across experiments:
- Double-nested response handling — the API returns
response.response, requiring a dedicated helper to unwrap cleanly - Retry with exponential backoff — AI inference can be slow; robust error handling was essential
- Structured prompt engineering — each experiment required carefully designed system prompts to get consistently parseable output
WordPress Plugin: AI Chatbot
The Lumination Chatbot is a WordPress plugin that adds a floating AI assistant to any website. It is currently running in production on a live site.
How It Works
A small chat bubble appears in the corner of every page. When a visitor clicks it, a conversation panel opens. The first message triggers the first API call — zero overhead until the user engages.
What makes this chatbot different from generic AI widgets is page context awareness. The plugin extracts the current page’s content server-side (stripping navigation, scripts, and footer), then passes it to the API as context. The chatbot can answer questions about the specific page the visitor is reading.
Admin Configuration
Site owners configure the chatbot through the WordPress admin panel:
- API key (stored securely, never exposed to the frontend)
- Bot name, welcome message, and placeholder text
- Primary colour to match site branding
- Custom instructions appended to every request
Built-in Analytics
Every conversation is tracked in a custom database table — timestamps, page URLs, session identifiers, token usage, and credits consumed. An analytics dashboard in the WordPress admin shows usage trends over time, top pages by engagement, and cost breakdowns by date range.
WordPress Plugin: AI Homework Helper
The Lumination Homework Helper takes the API further into education. This plugin lets students upload a photo of their homework — or type a problem — and receive a step-by-step, AI-generated solution with full mathematical notation.
Live Demo
Try it yourself — upload an image of a maths problem or type one in, and see the AI work through the solution step by step.
The Pipeline: From Photo to Solution
The plugin chains two API endpoints together:
- Upload & Solve — the student’s image (photo, screenshot, or PDF) is sent as base64 to the extraction endpoint, which identifies and transcribes the problem
- Chat Agent — the extracted text is wrapped in a structured educational prompt and sent to the conversational endpoint, which generates a detailed, step-by-step explanation
The result is rendered with Marked.js for markdown formatting, DOMPurify for security, and MathJax 3 for LaTeX mathematical notation — all bundled locally within the plugin, with no external CDN dependencies.
The LaTeX Challenge
Rendering AI-generated mathematical content in a WordPress environment required solving a specific technical problem: markdown parsers break LaTeX delimiters. The solution was a protect → parse → restore pattern — LaTeX expressions are extracted and replaced with placeholders before markdown parsing, then restored afterwards for MathJax to typeset.
This pattern, along with conflict detection (checking whether MathJax is already loaded by the theme or another plugin), ensures the homework helper works reliably across different WordPress installations.
Security and Compliance
The plugin follows WordPress Plugin Directory standards:
- API keys stored server-side, never exposed to the browser
- AJAX nonce verification on every request
- Rate limiting (10 requests per minute per user)
- File validation — MIME type checking, 10MB size limits, base64 sanitisation
- Full Privacy API implementation for GDPR compliance
The Vibecoding Advantage
Every product on this page was built through vibecoding — working with AI coding assistants to move from idea to functional product in hours rather than weeks. But the methodology goes beyond speed.
Rapid Validation
The chatbot went from concept to production deployment in a single working session. The homework helper followed. Each experiment on ai.azvai.com was built, tested, and published independently. This pace of iteration means ideas get validated (or discarded) quickly, with real users providing feedback on real products.
API as Creative Constraint
Working with a single API forced creative problem-solving. The same chat endpoint powers a customer service bot, a leadership coach, and a homework tutor — the difference is entirely in the prompt engineering, context handling, and frontend experience. This constraint led to reusable patterns: analytics tracking, rate limiting, response unwrapping, and error handling that transferred directly from one project to the next.
From Experiment to Product
The progression from the Azvai Experiments site to production WordPress plugins demonstrates a repeatable path: prototype quickly, test with real users, extract what works, and package it for distribution. The chatbot is live on a client site. The homework helper is ready for the WordPress Plugin Directory. The training tools are accessible to anyone at ai.azvai.com.
What Comes Next
The Lumination API continues to evolve, and so do the experiments. Planned next steps include:
- Moodle integration — bringing the homework helper and AI tutoring into learning management systems for schools and universities
- Streaming responses — real-time token-by-token output for a more responsive chat experience
- New training modules — expanding the experiments site with additional interactive scenarios
- WordPress Plugin Directory submission — making the homework helper freely available to any WordPress site
Want to explore what’s possible with AI APIs and vibecoding? Get in touch to discuss how rapid AI prototyping can solve real problems in your organisation.