My source
COROS watch
Excellent hardware. No easy public API for the history I wanted to explore.
From closed APIs to a personal fitness agent
COROS → agent
01 / Outcome
From fitness
data to
feedback
Now my fitness data explains itself every Sunday.
One data foundation. Many interfaces. Increasingly useful feedback loops.
Open the deckScan or tap
02 / Access before insight
The first problem was not AI.
It was getting the data out.
A great wearable is not automatically a useful data platform.
API
The obvious bridge
Direct Strava access
I did not want this personal project tied to another paid or managed API dependency.
The useful reframing
“Stop integrating every device. Find one trustworthy place for the history.”
COROS
Oura
other sources
→
one normalized layer
activities + recovery + history
The breakthrough was not a perfect connector. It was an abstraction layer I could keep reusing.
03 / Automation
First, I automated the bookkeeping
Airtable proved the ritual—and exposed the next gap.
Stravaactivity event
→
Zapiermove it
→
Airtableorganize it
→
Sunday emailreflect
airtable.com / fitness
It solved collection
Activities arrived consistently without manual bookkeeping.
It proved the ritual
A calm Sunday reflection was worth keeping.
The scope widened
Could one platform combine my COROS training, my partner’s Oura recovery, and new questions?
04 / Data foundation
The compounding decision:
one place for the data
Intervals.icu synced COROS training and Oura wellness, exposed an open API, and kept the core genuinely free.
intervals.icu / calendar
Why Intervals.icu
The platform I was looking for- Device syncsCOROS training + Oura wellness
- Activitiesruns, rides, swims, sessions
- Recoverysleep, HRV, resting HR, readiness
- Open APIactivities + wellness + history
- Two athletessame API shape, separate accounts
One data foundation. Many interfaces. Increasingly useful feedback loops.
05 / Exploratory analysis
Then I could ask questions
I had not planned for
I exported Intervals.icu data, uploaded it to regular Claude chat, and followed the questions wherever they went.
export→
upload→
ask→
insight
“Is my heart doing the same running for fewer beats?”
Claude turned exported rows into a chart and a plain-English interpretation.
Export, upload, prompt—and remember to repeat it.
Discovery was solved. Repeatability was not.
06 / Paperdink
The device is the scheduler.
SVG is the source of truth.
Paperdink puts my data beside my partner’s on one 4.2-inch e-ink display.
Installed at home
Prototype close-up
Kaushik
PERSONAL FEEDactivities + wellnessKathy
PERSONAL FEEDactivities + wellness
device drives cadenceno cronno database
07 / Purpose-built agents
The agent did not need
every agent feature
The SDK made the optional AI step easy. The useful product loop was everything around it.
General-purpose platform
OpenClaw-style breadth
many channelstoolsskillsmemorypluginsmodel routing
Useful when the job really needs a broad, programmable assistant.
choose the smallest runtime that fits→
This job
One narrow weekly loop
TelegramIntervals5 signalsSundaySQLiteoptional prose
Permissions, state, failure modes, and maintenance stay legible.
The SDK shelf is already here
GitHub Copilot SDK
Vercel AI SDK
OpenAI Agents SDK
Claude Agent SDK
“Connecting AI to a data warehouse is cheap. Deciding what it may do is the design.”
What I actually built
Custom Go orchestration + one SQLite file + Telegram. The Copilot SDK is one fresh, tool-disabled, optional narration call.
No dedicated hosted backend: it runs on hardware I already own and uses existing service accounts.
08 / Fitness data, translated
Five signals become
five human questions
The report compares me with my own recent normal—usually six trailing weeks—not with another athlete.
Load & ramp
Is my longer-term fitness trend rising, holding, or easing?
- CTL
- Chronic Training Load · rolling ~42-day average; a proxy for longer-term fitness
- ATL
- Acute Training Load · rolling ~7-day average; a proxy for recent fatigue
- TSB / form
- Training Stress Balance · CTL minus ATL; positive is fresher, negative means carrying fatigue
- CTL ramp
- week-over-week CTL change; shows whether fitness is building or fading
Volume
How much training did I actually do?
- Moving time
- minutes actually moving; pauses do not count
- Training load
- Intervals.icu stress score for one workout; weekly load adds them together
Intensity
How hard and concentrated was the work?
- Load density
- weekly load divided by moving hours; how concentrated the stress was
- Pace
- median running minutes per mile, with implausible values filtered out
Consistency
How often did I show up?
- Active days
- days with at least one training session
- Sessions
- number of completed workouts in the week
Recovery
Do readiness signals support the training story?
- HRV
- Heart-rate variability · rising versus your normal usually means better recovered
- Resting HR
- resting heart rate · falling versus your normal usually means better recovered
- Sleep
- average nightly sleep duration for the week, when available
How it stays honest
Medians soften one strange week. Missing inputs lower confidence and add a caveat—they never become a guess.
buildingholdingrecovering by designdecline possibleoverreaching cautioninsufficient data
09 / Architecture + trust
The model never
invents a number
AI writes above the report. It never calculates a metric, changes a verdict, or controls the schedule.
Optional sidecar
GitHub Copilot SDK finished facts only · tools off · 1–2 sentence lede ↓
Telegram intent
→
on plan · deliberately easy · off plan
→
classifier
Sunday schedule24h catch-upstale data labeledWAL-safe backupsbuild / backup failures alert
10 / The product surface
Telegram is
the frontend
Live demo
Buttons, commands, and saved messages remove the need to build a separate app UI.
What the finished report explains
- 01Verdict + weekone combined read and why
- 02Actual sessionssport, time, distance, pace, HR
- 03Five signalscompared with recent personal normal
- 04Caveats + takeawaywhat is missing and one gentle next thought
Jun 29–Jul 5 Building
Jun 22–28 Recovering by design
Jun 15–21 Holding
/report [date]- build a week
/checkin [date]- intent + report
/history- 10 saved verdicts
/last- resend newest
/signals- plain glossary
/help- the whole UI
Demo path
/checkin a completed week→point to verdict + sessions + baseline→/history
11 / Build with confidence
Make every layer
inspectable
Helpful commands turn an agent from a magic demo into a system you can understand, test, and extend.
make dump-week DAY=...see raw weekly aggregatesmake facts DAY=...inspect five signals + verdict JSONmake report-rawpreview deterministic Telegram outputmake reportcompare optional AI narrationmake verifyformat, vet, test, build, glossary
make preview-p1open the exact SVG in Chrome/previewsee both pages + simulated 1-bit modemake curl-bmpverify 400 × 300 bitmap encodingmake typecheckcatch Worker contract mistakesmake flashship only after the browser path is right
Do not debug the whole agent at once. Make the source, facts, render, and delivery independently visible.
12 / Next stop
More context.
Same small loop.
Once the data has a home, one more signal or command is an extension—not a new product.
trusted personal data
shared facts + history
- 01Centralize firstReliable history creates optionality.
- 02Keep facts inspectableAI can explain without owning the truth.
- 03Build the smallest useful loopOne purpose, one interface, clear failure modes.