Harmony Pulse

Problem

A rider gets one or two coached lessons a week and rides alone the rest of the time. Those solo rides have no feedback at all. Nothing speaks up in the moment, and afterwards the only record is how it felt.

Harmony Pulse gives those rides a voice and a record. Two chest straps, one on the horse’s girth and one on the rider, feed an iPhone in her pocket. During the ride it speaks short cues about rhythm, breathing and how hard the horse is working. Afterwards it grades the session, writes a debrief, and produces a one-file report her trainer can read before the next lesson.

The rider is my mom, Kirstin. She has two horses, Bowie and Denali, and she’s the domain expert and the first tester. The showcase page is the version she shows people. This page covers how it’s built.

Constraints

• Ranches have terrible wifi and worse cell service. So nothing in the app can need the internet. Recording, coaching, grading and the AI debrief all run on the phone, and there’s no account, no cloud and no subscription.
• No commercial horse sensor will share its data. Every one I looked at is a closed ecosystem with no API. The horse wears the same off-the-shelf Polar H10 the rider does, talked to over plain Bluetooth with no vendor SDK, and its accelerometer on the girth is the gait sensor.
• It talks into someone’s ear while she’s on a horse. A cue that’s late, wrong or chatty is worse than silence, so the live path has to be boring and predictable.
• Two heart straps can’t diagnose anything. The app compares a horse only to his own recent history and never names a limb. That rule is written down and tested.

How it works

I built the recorder first and the analysis second, and most of the design follows from that order.

Every byte from every sensor is written to disk as it arrives, before anything decodes it. A recorded ride and a live ride then go through the same code. Below one seam the app can’t tell a real strap from a replayed file or a simulated horse, and the analysis is deterministic, with no clock reads and no randomness. Replay a ride twice and the event log matches byte for byte. A decoder fix therefore repairs every ride ever recorded, and the whole pipeline, from sensor frames to spoken cues to the trainer report, runs at a desk with no hardware and no horse.

The coach sits on top of that. Ten times a second it updates what it knows about gait, stride rhythm, transitions, the horse’s effort and recovery, and the rider’s breathing and heart rhythm. Detectors turn changes into events, and a cue policy decides which events are worth her attention. Most aren’t. The sample ride on the showcase noticed 28 moments and spoke about six. The words come from a phrase bank written by hand and recorded in two voices, and the recordings ship inside the app.

AI is used only after the ride, and only to write prose. The session is summarised into numbers in Swift, Apple Intelligence rewrites those numbers as a debrief on the device, and a deterministic honesty check reads the result before she does. If the model isn’t available, or it says something the numbers don’t support, she gets a plain structured debrief built from the same numbers.

What I’ve built

• A Swift package that holds all the logic and depends only on Foundation, so it builds and tests on Linux in CI. It has more than 500 tests, including a golden event log that fails the build when a change moves a single cue.
• The iPhone app in SwiftUI: pairing, the live ride screen, the report card, heart-rate charts, a route map, session history, and a trainer report exported as one HTML file.
• An Apple Watch app that owns the workout session and gives her a Log button and press-and-hold voice notes, so she doesn’t have to reach for the phone.
• Quests, which are short scripted sessions where the coach tells her what to do and when. She gets a guided exercise and I get labelled data, so a change to the gait or breath detection can finally be judged by a number.
• A set of decision records. A lot of the interesting choices were things I decided to leave out.

Current state

It’s in testing with one rider and two horses. She sends real sessions from the barn, and the work has moved from building to tuning, which means checking that what the app reads is what happened.

Gait at walk and trot was retuned on real strap data, and the rider’s calm reading was checked against an open breathing dataset. Other readings aren’t there yet. Canter has the least real data behind it, transitions are unvalidated, and the breath-hold detector flags far too often, so that one gets fixed first.

There’s no launch date and no business plan. The goal is to make it everything she wants it to be. Her next idea is a coach that keeps what her trainer said in a lesson and turns it into a plan for the week.