Building in public.
Truth over marketing
We don't invent numbers or hide our flaws. You see the engineering exactly as it happens.
Silence is golden
We don't overwhelm you with raw data. We only alert you when something actually drifts from your dog's baseline.
No sponsored bias
We don't sell your data or push paid pet food. Our tools rely entirely on verifiable facts.
November 2026
Web Summit: everyone talked about AI. The health section made sense of it
At Web Summit in Lisbon almost every stand was about AI agents, SaaS and enterprise. Our stand was in the health section, and that is where AI finally made sense.
A little cow data almost broke our dog model
We went looking for more chewing data and found it in cattle research. A small slice of it made our dog model worse overall. Why the right data is gold.
October 2026
Updating our server and collar without downtime
How we made server deploys swap blue for green with no gap, how the collar now updates over Bluetooth from the app, and why a prototype collar can't do blue/green yet.
Making our collar listen only to its owner
Our prototype collar obeyed any phone in Bluetooth range. Why ordinary pairing could not fix that on a collar with no buttons, and the per-collar key and signed commands we built instead.
Fixing collar sync and battery life, and reaching Google Play
How a firmware audit, a flash log and two sync fixes took our prototype collar from a battery that always read 100% to a measured 28 days on a small test cell, and how the Android app reached Google Play.
Rebuilding fudini.co around the collar
Why the site now leads with the collar instead of the app, and the claims, renders and stylesheet we had to fix on the way.
June 2026
The LED that ate our battery
How status LEDs drained our prototype collar in 7-10 days against a power budget of 75-100 days, and why the first fixes missed.
Our AI called 41 °C a good day for a walk
Why you should never trust an LLM to evaluate numeric thresholds, and how we fixed our walk-weather widget.
April 2026
Teaching a neural net what a dog eating looks like
How we calibrated synthetic IMU signals with allometric scaling and head-bob physics to detect pet feeding.
Why we stopped using Zephyr on the XIAO nRF52840
Zephyr crashed on the XIAO nRF52840 Sense's UF2 bootloader, so we moved the collar to Arduino NimBLE. What broke along the way.
March 2026
Native iOS background Bluetooth sync
Moving collar BLE sync into a native Swift CoreBluetooth module, then replacing timers with a BLE heartbeat to survive iOS background suspension.
A fixed sampling interval saw 3.3% of the day
Why we replaced a fixed 60-second sampling interval with adaptive sampling, and cut BLE syncs from about 60 to 12 an hour.
Moving activity classification off the collar
We stripped on-device variance classification from our collar firmware and moved activity classification, then a neural net, to the backend.
Explore the hardware.
See the specs and timeline for the collar. Built for the collar they already wear.
Subscribe to
updates.
Get new logs in your inbox. Free to join.