Android On-Device Animal Identification in Practice: From Training to LiteRT Inference How Safari Go moves from PyTorch training and LiteRT export to local animal identification on Android, with observed GPU probing and a deterministic CPU fallback.
Read article
Field notes
What changes when the model lives close to the work.
We write about on-device models, mobile and field work, private data, offline behavior, and the small product decisions that decide whether an AI tool survives its first week.
Android On-Device Animal Identification in Practice: From Training to LiteRT Inference How Safari Go moves from PyTorch training and LiteRT export to local animal identification on Android, with observed GPU probing and a deterministic CPU fallback.
Read article
The writing tool started with my wife My wife wanted help getting back into writing. I thought GPT would make the tool easy. It did not work out quite that cleanly.
Read article
The first private knowledge system an SMB should build The first useful private knowledge system should answer repeated operational questions from approved sources and make its evidence visible.
Read article
Most SMBs do not need an AI strategy. They need one broken workflow fixed. The useful starting point for SMB AI adoption is not a strategy deck. It is one repeated workflow that costs time every week and can be improved with a small system.
Read article
When cloud-only AI is the wrong deployment model Cloud AI is often the easiest starting point, but some SMB workflows need local capture, on-device models, private networking, or hardware deployed near the work.
Read article
Why Plant ID Is Not Just a Taxonomy Problem Plant identification looks solved on paper. Real-world plant ID is a noisy visual classification problem, not a taxonomy traversal. Here is what we learned building PlantCare.
Read article
Stop rebuilding the planner every time a tool changes Most agentic systems we inherit do the same thing: they treat the planner as glue between tools, so changing one tool breaks the planner. Here is the layout we keep coming back to.
Read article
Sensible defaults for a private RAG that you can actually maintain Most private RAG projects die from operational debt, not from retrieval quality. These are the defaults we set on day one so the system stays maintainable past month three.
Read article
The minimum useful observability for an LLM-backed service You do not need a dashboard. You need three signals, written down on day one, that tell you whether the system is doing what you expect.
Read article
From note to system
Have one workflow that looks like this?
Bring the repeated loop, the current tools, and the documents people already trust. We will map the first useful AI system around it.