An AI coach that knows my constraints

Accruing data is easy.
Using it is hard.

Tracking apps log what already happened. Training plans assume a schedule and a life I don't have. So I built something that did. An AI running coach that uses my daily biometric data to keep marathon training on track through everything a training plan doesn't consider.

Top results

+0% HRV balance · pre vs post Based on pre vs post coach system HRV data. Autonomic recovery improved with adaptive monitoring and plan adjustment.

0% Plan adherence · 12 weeks 367 miles planned · 362.2 actual.

0 Consecutive daily briefings One briefing every day, Jan 24 through Apr 19.

Multiple integrations.
One Frontier Model.

A scheduled Cloud function pulls yesterday's sleep, HRV, and readiness from Oura; yesterday's runs from Strava; weight from Withings; the next two weeks of clinical shifts from Google Calendar; and the morning's conditions from a weather API. Each source is provided by a custom-coded MCP server so the orchestrator calls them the same way.

It then pulls the activity, pace, and coach history from Firestore, every metric, every piece of advice the system gave, and the current active plan, and assembles a single prompt: biodata, calendar, active plan, strategy rules, recent history.

That prompt goes to one frontier model. The model alters the plan when needed, stores necessary memories, and sends the briefing. The briefing lands on the phone through Telegram. 86 days, 86 briefings, so far.

The model optimizes against rules the user sets.

The coach is initialized to know what post-call means. The system follows user-defined constraints and works to balance those against realities of scheduling, weather, and biometrics.

Example 01 · OR service

Day-shift OR rotation
(7a – 7p)

→ training moves to evening

If the user is working all day, the coach adapts the plan. Lifts and easy runs land after 7 PM; nothing before.

Example 02 · Sequential nights

Nights Shifts

→ mandatory rest block

It knows shift starts and ends and that time between night shifts is not conducive to hard training

Example 03 · Post-call

Morning after a call shift

→ How fast can things resume?

Pulls HRV on recovery night to determine degree of recovery needed before restarting the plan

Monday, March 2Closing Week 5 — Sequential Night Shifts

Calendar · the constraints the model saw

Mon · Mar 2 · 2026 WK 5 · D7

Recovery — sleep till noon post-call · protected

Research check-in 2 PM · 30 min

Active recovery — easy 4 mi RPE 3 · preserve Monday

Telegram · 6:04 AM

Coach

daily briefing

Status: Yellow Light (Post-Call). You just finished two sequential night shifts. Readiness sits at 62 — below the 75 threshold the plan respects for post-call training. Your autonomic system is still settling. Pushing a quality session today would cost more than it buys.

The Adjustment. Swapping today's track workout for Recovery 4 mi, Zone 2 only. Cap it at RPE 3/10 — if HR drifts above 145, walk it in. Tuesday's lower-body lift holds. The long run stays on Saturday — that's the anchor.

read · 6:04

HRV improved under a guided schedule.

Across 149 days of continuous Oura data, every recovery metric moved in the right direction. The best — HRV increase of 28.5% — is the signal I care about most.

HRV balance · pre vs post coach

+28.5%

An 18.4 ms rise in Oura HRV balance — sustained across three months.

81 days pre 68 days post 149 measured

Recovery & readiness

Every Oura recovery increased.

Adaptations across 86 coach runs

Weekly miles · planned vs actual

98.7% adherence across 12 weeks

362.2 mi actual · 367 planned

Actual Planned

Weekly mileage rises through a training block and tapers before a goal race. The system chose the day-by-day shape; this is the cumulative result.

Accruing data is easy.Using it is hard.

The problem

Multiple integrations. One Frontier Model.