Product partnershipConsumer product

Rutz — an AI running coach

A product I’m building with a partner: Garmin data becomes running analysis, with calculations in code and explanations from AI.

Run receivedRaw Garmin data
ScoredCode: same run, same numbers
ExplainedAI: analysis in plain language
Maya decidesWhat to do with the advice

Code computeddeterministic

Run8.2 km · 42:10

Score7.4 / 10

Training load68 · +12%

DuplicateNone

Pattern3rd hard day in a row

AI explained06:58

Three hard days in a row and your weekly load is up 12%. Tomorrow, keep it easy: 30 minutes at a conversational pace.

Why 30 minutes?

Analysing…Maya choosesFictional data

Recreated interface scene with fictional data

The product

Rutz is an iOS app I’m building with a partner. It turns runners’ Garmin data into analysis they can understand and act on. Code handles the calculations, AI helps interpret the results and the runner decides what to do with the advice.

The design question

How do you turn running data into advice a runner can understand and assess? I designed a separation between reproducible calculations and language-model interpretation, so the explanation is grounded in computed results.

What was built

Code does the math: Scoring, de-duplication, training-load math and pattern detection run as deterministic code. The same run always gets the same numbers.
AI interprets the results: The model gets computed results, not raw data, and explains what they mean: a heavier post-run analysis and a lighter chat.
Garmin data pipeline: Each runner's Garmin activity flows through their own connection into Postgres, where the analysis code picks it up.
iOS app and backend: A SwiftUI app with Apple Sign-In, a FastAPI backend with plain parameterised SQL on Postgres, deployed on Railway.
Internal run-analysis tooling: A developer tool that replays runs with changed parameters in real time, so a scoring change can be seen before it ships.

My part

I’m responsible for Rutz’s architecture and AI design: how the systems connect, what is calculated in code and what the model receives to interpret. My partner wrote most of the code.

Supported outcomes

AI is used only where interpretation is needed
Every number in the app is reproducible from code
The runner stays in control of what to do with the advice

Tools

SwiftUI, FastAPI, Postgres, Garmin, Claude, Gemini, Railway