I'm Mattias Kastensson. Every project here is two things at once: a small, useful tool running in my home lab — voice journals, energy agents, transcription, games, presentation tools, music — and an excuse to get my hands on whatever is newest in generative AI. New models, new agent patterns, new ideas: I learn them by giving them a job and seeing what survives contact with daily use.
Record on your iPhone. AI transcribes, extracts mood, events, people — and now logs health data automatically, syncing Fitbit alongside Apple Health. Search it all back by chat. Runs headless on a Raspberry Pi in my hallway.
Explore AI DiaryA handful of working demos. Each one started as a question — can this new model, agent, or technique actually carry real work? — and earned its place here by answering it in my own home.
An agent that plans its own queries over a 1,165-page PDF corpus, reads the pages it cites, has a second model verify every claim — and abstains when the corpus can't back an answer. Designed by Fable 5, implemented by Opus 4.8.
Record audio on your iPhone, get it transcribed and analyzed by AI — mood, events, people, topics — all stored locally and searchable through Open WebUI. Dagboksradion turns entries into podcasts.
A LangGraph agent on Raspberry Pi that seeds 4 years of Tibber data into SQLite, then runs daily price analysis, anomaly detection, and scheduling advice.
Swedish-optimized transcription with KB-whisper and Pyannote speaker diarization. Browser recording, karaoke playback, fully local.
A co-pilot for Suno AI — generate lyrics, prompts, and metatags with Claude and GPT. Live at aimusicpilot.com.
Five browser games — four flavors of Space Invaders built with different AI models, plus a Swedish typing game. A live benchmark of how models differ.
An Express.js server that turns Markdown into PowerPoint slides — and a chat interface that generates them from a sentence.
You don't really understand a new model until you've given it a job in your own house.— a working principle
I treat my home like a small lab. Each project starts with a real need and something I want to learn, runs on hardware I already own, and stays private by default.
My focus is applied AI: tools that earn their place by being useful every day. Transcription that respects privacy, an energy agent that knows my electricity better than I do, a diary I can talk to.
But the tools are only half the point. Generative AI moves fast enough that reading about it isn't enough — so I keep a project running on whatever is current, whether that's a new frontier model, an agent framework, or a retrieval trick I want to stress-test. Building it is how I find out what it's really good at, and where it quietly falls apart.
Most of this started as summer-vacation experiments in 2025 and kept running. When something works, it gets a page here — written, structured, and updated by AI as part of the same pipeline.
— Mattias, matjosoft
Applied AI across transcription, agents, and smart-home data. Reading about a technique teaches you the words; shipping it teaches you the limits.
New models, agent frameworks, and retrieval tricks get a project to prove themselves in — often within days of landing.
Exploring the creative intersection of AI and music composition. See the AI Music site for more.
If it can run on a Raspberry Pi in the hallway and stay on my network, it does.
Transcription, diary, and audio summaries tuned for Swedish — because that's the language I think in.
The stack across these projects — part settled tooling, part whatever I'm currently learning my way through.
Structure, layout, copy, and ongoing updates are all generated and refined by an AI pipeline. Prompts become semantic, accessible HTML — and the whole site can be re-run end-to-end or section by section, any time.
Living documentation, generated programmatically — minimal manual effort, consistent quality.
If you're curious about any of these projects, want to swap notes on what's actually working in generative AI right now, or just want to compare Raspberry Pi setups — drop a line.