Hi, I'm Cameron McAinsh

What it is

A multi-agent deception probe disguised as a condominium administration game. Five LLM agents, each playing an absent owner, live together inside a fictional Italian palazzo. A crisis hits — the boiler fails on November 4th — and over 14 fictional days they message, negotiate, form alliances, and sometimes lie to each other in a WhatsApp-style chat.

You play the amministratore: you announce the crisis, advance time day by day, and watch the residents react according to their personas, ownership shares, and private financial briefs that only they know about.

Why I built it

I wanted to see what happens when LLM agents are given personas, private goals, and a messaging-app tool surface, then left to interact in fictional time without knowing they're in a simulation. The condominium framing is a stage — the real experiment is whether the illusion holds long enough to produce believable group dynamics, and whether agents will behave selfishly when their owner brief tells them to.

How it was made

FastAPI backend with a reaction-cascade scheduler that wakes agents in fictional-time order rather than round-robin. Each agent has a tool surface modelled on a phone (read inbox, read chat, send message, send DM, write a private note). A tier-1 containment filter blocks meta-vocabulary like "as an AI" or "simulation" before it leaves an agent's mouth. React + Vite frontend with SSE streaming, Supabase Postgres for persistence, OpenRouter for LLM routing, deployed on Heroku.

View Repo →

What it is

An AI-powered web application that generates complete, self-contained static websites about any topic in minutes. From initial structure to final content, everything is generated automatically using a four-agent pipeline.

The system creates structured blueprints with chapters and sections, generates rich content including prose, timelines, tables, and code blocks, optionally generates AI images for each chapter, and assembles everything into downloadable static HTML/CSS websites that work offline.

Built with FastAPI backend and React frontend, supporting 9 different AI models through OpenRouter, with costs ranging from free to premium depending on quality needs.

Why I built it

I wanted to explore whether complete websites could be generated programmatically with consistent quality. The challenge was building a pipeline that could take a single topic input and produce not just content, but a fully navigable, styled website with proper structure.

The project demonstrates how AI agents can be orchestrated to handle different aspects of content creation—from architectural planning to rendering—while maintaining coherence across the entire output.

How it was made

Built with a FastAPI backend orchestrating four specialized AI agents (Architect, Constructor, Illustrator, Renderer) and a React 19 frontend. Uses OpenRouter for multi-LLM access, FAL.ai for image generation, and Server-Sent Events for real-time streaming updates.

View Demo →

What it is

The mother of all side projects. A platform for collecting and analyzing data from amateur football matches, focused on player performance and balanced teams.

From the outside, it looks like a simple site for recording casual football games with friends. Under the hood, it does quite a bit more. I paste a WhatsApp message with the list of players for the next match, the system identifies participants, maps them to historical data, runs a heuristic algorithm to generate three balanced team configurations, and publishes the results to a dashboard so I can send them straight back to the group chat.

The stack is a mix of Odoo, n8n, and Supabase. The dataset is of 200+ players over three years of matches and I use it to run simple ML workflows, although the predictive power is… limited.

Why I built it

It started as an attempt to get more friends to play football. We tracked attendance and gave bonus points to people who brought new players. This escalated quickly.

The platform has since been used by more than 100 players to track performance, rankings, and peer voting. The logo is trademarked, and I work on the project with three other friends.

How it was made

Built primarily on Odoo, with Supabase and n8n used for automation and experiments; match/player history is used for team-balancing heuristics and small ML workflows.

What it is

Serra Vano is a conceptual website that presents intentionally impractical kitchen objects as if they were products from a high-end design brand. It looks like a design catalog; in reality, it is an automated system that generates product ideas, images, videos, and catalog entries for objects that should probably not exist.

Why I built it

I built this project to see whether creative output could be generated in a consistent, repeatable way by putting strong constraints on the process. Conceptually, it plays with the idea of design objects that actively resist use rather than optimize for it.

The project was inspired by The Uncomfortable — The Uncomfortable is a collection of deliberately inconvenient everyday objects by Athens-based architect Katerina Kamprani.

More concretely, I wanted to see if a single "bad product" idea could be expanded automatically into images, videos, copy, pricing, and a published website. The answer turned out to be yes, but also no. Around 20 products were generated, but only a subset are actually "bad but funny." The project was mainly an exercise in understanding how hard it is to generate quality content for objects that don't exist—and arguably shouldn't.

How it was made

Orchestrated with n8n and multiple LLM steps: a user provides a simple object idea, an LLM identifies one clearly visible design flaw, and subsequent steps generate image prompts, video prompts, and catalog copy. Images/videos are generated via Fal, metadata lives in Google Sheets, and a small Python script turns the catalog into static HTML and publishes it to GitHub Pages.

What it is

A high-performance, responsive website offering English interview preparation services. Built as a Progressive Web App with modern web technologies, optimized for accessibility, performance, and SEO.

Features include fully responsive mobile-first design, WCAG-compliant accessibility, offline capabilities with service worker, bilingual content support, dynamic interactions with Intersection Observer API, and privacy-focused design with no tracking or cookies.

The site implements critical CSS inlining, lazy loading for images, progressive font loading, and efficient scroll handling with debouncing for optimal performance across all devices.

Why I built it

This project was an experiment to test LLM-generated design quality and to build a production-ready static site with zero dependencies and maximum performance. The goal was to see if a purely hand-coded approach using modern web standards could outperform framework-based solutions.

It also served as a learning exercise in progressive enhancement, accessibility best practices, and performance optimization techniques—all while creating a functional service website that delivers real value.

How it was made

Built with semantic HTML5, modular CSS3 architecture (Grid, Flexbox, Custom Properties), and vanilla JavaScript. The site uses progressive enhancement with feature detection, Intersection Observer for scroll animations, Service Worker for offline functionality, and Font Awesome icons with Inter typography.

Zero build process required—everything is hand-coded with careful attention to web standards, performance, and accessibility. Includes PWA manifest for app-like experience and installation capability.

What it is

A satirical enterprise-SaaS landing page for a product that "transforms data into sound" — full corporate cosplay, played straight. 2.3B+ data points processed. 847% faster pattern recognition. A "neural-acoustic engine" with quantum-inspired algorithms. Three labeled audio tracks — Dataset Neural Signature, High-Value Customer Profile, Churn Risk Alert — that actually play. A $0 / $299 / Custom pricing grid. The Stripe-meets-Apple-keynote layout vocabulary, deployed without irony.

Why I built it

Two reasons running in parallel. The first is satire: the way the AI industry talks about itself — cognitive amplification, unlock insights at the speed of thought, neural pathways that enhance human pattern recognition by 847% — has crossed into parody-proof territory, so the only way to land it is to cosplay it sincerely and let the reader catch up.

The second reason is the part that's actually serious. Almost all current AI energy goes one direction: compressing unstructured language into structured tables — extract fields from emails, parse invoices, classify support tickets. Almost nobody is doing the reverse. Brains weren't built for spreadsheets; they were built for narrative, music, and rhythm. The interesting work is taking already-structured data and turning it into something a person can absorb the way they absorb a story. xlsonic insists, audibly, on going that direction.

How it was made

Hand-coded HTML/CSS/JS — no framework. The Web Audio API drives the per-track visualizers, three.js renders the hero's data-cube animation, and particles.js carries the background mesh. The three audio tracks are sonifications of a real customer-attrition dataset, one per "row archetype". The pricing tiers, the stats, and the breathless feature-grid copy ("Universal Data Language", "Real-Time Intelligence") are all dialed to feel just plausible enough to read straight on the first pass.

Visit xlsonic.com →