Personal AI platform
Hearth
A complete personal AI platform running on my own hardware — gateway service, local models, an agentic loop, voice, vision, and a native Android client. The SMS categoriser that started it's now one feature out of dozens.
- Home
- The Lab
The Lab
What I build when
nobody is paying me
The client work on this site is WordPress, because WordPress is usually the right answer for a service business. It's not the extent of what I do, and a portfolio of Divi builds is a poor description of an engineer.
So this is the other half: Python services, native Android written against the framework rather than a toolkit, local language models running on my own hardware, containers, MCP servers, a game engine. Built for me, on my own time, because I wanted the thing to exist.
The two big ones
Where most of the hours went
One is a personal AI platform I've been building for over a year. The other is the plugin and API that lets an assistant work on a live WordPress site — and the reason the client work on this site can move as fast as it does.
WordPress plugin + REST API
WP AI Agent
A WordPress plugin and custom REST API that lets an AI assistant work on a live site directly — read pages, build them, publish them — with authentication, a tiered permission model and an audit log. Around 5,900 lines of PHP across 15 files, paired with a deliberately thin Node MCP server.
Local AI infrastructure
A private AI platform running entirely on hardware I own. No API keys, no per-token bill, nothing leaving the network.
This site's assistant
The assistant on this site runs on a GPU in my office, answering from the pages you're reading. It is the only local model I run that strangers can talk to, and in September somebody spent forty minutes trying to break it. The write-up is what they got through and what finally stopped them.
Running models locally
What a parameter is, why a 7B model is 14 GB before anyone compresses it, and how to work out what your own graphics card can run before downloading twenty gigabytes to find out. Written for someone who has a gaming card and has not tried this yet.
claude-local
Runs Claude Code against a model served locally by LM Studio, with no hosted API in the loop. No key, no per-token cost, and no source code leaving the machine — which matters when the code belongs to a client.
AI Hub
A containerised dashboard for driving local models, sitting on Docker with Ollama on the host. The control surface for everything else on this list.
vram-toggle
Two batch files that free roughly 12 GB of VRAM on an RTX 4090 by pausing the services holding it, unloading anything resident, and putting it all back afterwards. Unglamorous, and I use it constantly.
Local Voice Assistant
A fully local voice assistant on a Raspberry Pi with a small display, talking to the same gateway everything else uses. Nothing it hears leaves the LAN.
Wake-word training
Training a custom "hey Hearth" wake word from scratch — building the dataset, mixing in room impulse responses and background audio, and evaluating the result.
Local image generation
A Krea 2 Turbo checkpoint running in ComfyUI on the same machine as the rest of my local AI work, driven by a small Python client that turns a one-line brief into a finished, correctly-sized image in about eleven seconds. Built to keep diffusion models on hardware I own — no API key, no per-image bill, nothing leaving the network.
Native Android
Self-signed, sideloaded, no Play Store. Written to solve a specific problem on a specific phone, with no intention of shipping it to strangers.
Hearth (Android)
The native Android client for the gateway. Java, framework-only — no AndroidX — with a minimal command-line build. A WebView of the web app plus the native capabilities a browser can't reach.
Hearth SMS
A full replacement for the phone's default messaging app: SMS and MMS, group threads with stitched avatars, search across message text and contacts, gallery attachments auto-compressed to the carrier's real limit read at runtime.
Fitness coach
Built for my brother. Set a current weight, a target and a deadline; it builds a daily diet and workout plan, writes it into Google Calendar, and tracks calories and weight against the goal.
Developer tooling
The machinery behind the client work — MCP servers, generators, dashboards and CLIs that turn repeated jobs into one command.
Willits Forge Booking
An appointment booking plugin with two-way Google Calendar sync, written for practitioner sites. Its own database table, a custom REST namespace, configurable availability rules, transactional email over SMTP, and a self-service cancellation flow. Currently running on a live client site.
Divi 5 Generator
Answers to a short interview in, a complete on-brand Divi 5 site out — page map, brand, copy scaffold and one validated portability export per page.
MCP server collection
A set of Model Context Protocol servers that give an AI assistant real hands: WordPress sites, my phone, the television, image generation. Each one is a thin, well-scoped bridge and not a grab-bag of endpoints.
Fleet dashboard
Polls every site running the agent plugin and reports the fleet: online or not, plugin version drift, WordPress and PHP versions, builder fingerprint. Answers 'is everything up' without opening 30 tabs.
Experiments
Things built to find out whether they could be built.
Hero Line Wars
A mobile real-time strategy game in the Warcraft III custom-map tradition — auto-spawning units, army customization, hero-led battalions and bidirectional fog of war. Simulation foundation first, engine second.
LLM baseline harness
A repeatable harness for measuring local models against a fixed prompt set, so 'is the new model actually better' has an answer and not a vibe.
Why this is on a business website
Because the stack should follow the problem
A developer who only knows WordPress will recommend WordPress for everything, because it is the only tool in the bag. That's not advice, it's a reflex — and you can't tell the difference from the outside until the project is already in trouble.
Knowing what a Python service, a native app, a container or a queue actually costs to build and maintain is what makes “WordPress is the right call here” worth hearing. Most of the time it genuinely is: it wins on total cost of ownership and on content workflow, and I'll say so. When it's not, you get told that instead — and the recommendation comes from someone who has built the alternative, not just read about it.
None of the work on this page was commissioned. It exists because I wanted it to, which is also the honest answer to whether I enjoy this.
Got something that's not a website?
Automation, an internal tool, an integration, an app, a system that needs to talk to another system. Tell me what it's and you'll get an honest read on whether I'm the right person for it.