imessage · agentic · evals
iMessage Baby Tracker
the group chat → what got logged
thursday
Maya · 7:12 AM
Morning! 5oz bottle at 6:45, he crushed it
→ feed · 5 oz · bottle · 6:45 AM
Maya · 9:32 AM
down for a nap
→ sleep · started 9:32 AM
Sam · 10:15 AM
he's up
→ sleep · nap closed, 9:32–10:15
Rosa · 10:41 AM
le di 4 onzas a las 10:30, muy contento
→ feed · 4 oz · 10:30 AM — extracted from Spanish, fields normalized
Rosa · 11:22 AM
10 minutos de tummy time, no le gustó jajaja
→ tummy time · 10 min · mood captured
Maya · 3:20 PM
so cute 😍
→ skipped — chatter, not data
friday
Maya · 6:52 AM
up at 6:30, nursed 20 min
→ feed · 20 min · breast — never converted to ounces
Deb · 9:47 AM
FIRST time at the farmers market!! he loved the flowers
→ activity · special → the story page
Maya · 2:33 PM
weighed him at home — 14 lbs 2 oz!
→ measurement · 14.125 lbs
Rosa · 5:31 PM
5 onzas a las 5, y está un poco irritable esta tarde
→ feed · 5 oz + mood note · irritable
excerpt — the full fabricated day was 25 messages → 28 typed events, all classified by the real pipeline
the briefings texted back to the chat, 7:00 & 20:00 — verbatim
☀ morning
Yesterday (Thursday, September 10) he slept a solid ~10.5 hours overnight (630 min, from your reported bedtime — no sock synced, so no waking data to read). He napped twice for 160 minutes, and came across mostly happy with just one fussy stretch. Two spots to note: tummy time landed at 10 of the 30-min goal, and intake was only 15 oz across 3 feeds — well under the 28-oz target.
Intake is the thread to watch: yesterday 15 oz, today tracking similarly low (14.3 oz, partly estimated). At 19 weeks (~4 months) appetite dips and distractibility around feeds commonly coincide with the 4-month developmental leap and sleep regression window — often temporary. But two low-intake days in a row, if feeds keep dropping or he seems off, is worth mentioning to the pediatrician.
Today's practical focus: tummy time has been running short (10 then 15 min) — a couple more short floor sessions would close the gap to 30. Exercise sessions are also at 1 of 3 both days, so there's room for his neck work. Keep an eye on those feeds too.
☾ evening
Today's been a bright one for him — three happy notes and a content stretch, with just one fussy patch. He got 15 minutes of tummy time and one exercise session in, plus a solid 81-minute nap (that count is reliable). Feeds landed around 14.3 oz so far, though about 5 of those were estimated.
Still open before bath: he's 15 minutes shy of the 30-minute tummy time goal, and two exercise sessions short of three — so there's room for a gentle session or two if he's up for it, good for both the flat spot and the tight neck. Feeds are tracking under the 28 oz goal too, so evening ounces will help close the gap.
Setting up the night: last night was a smooth one — 10.5 hours reported with zero sustained wakings the sock caught, which is lovely at ~4 months. Around this age sleep can still shift as things develop, so don't read too much into any variation.
the dashboard, snapped midday — scroll it
the family is fabricated (Sam, Maya, Rosa, Deb) — everything else is the real system: real classifier on every message, briefings verbatim from the model, dashboard rendered from the resulting database.
What it does
I built this for my infant son. Every baby app makes you tap through dropdowns and input fields to log a feed. This one has no interface at all: four caregivers text the family group chat in natural language, in two languages — times and amounts of feeds, baths, naps, diapers, moods, and the events we want to remember, like the first time meeting so-and-so. Messages become typed events with daily goals and weekly trends, a dashboard shows the current picture, and morning and evening briefings go back to the group chat, reading the log and goals against his age to give useful guidance.
The biggest hassle of infant care is the transfer of information between caregivers — this answers that directly.
How it works
Group chat in → typed events out → goals, trends, and briefings on top. The classifier has its own eval suite (~48 cases, runs for about a cent), and any language works — one grandmother texts in Spanish. The rest stays private with the repo.
What broke + what I changed
The main thing that breaks: it's iMessage-based, so it depends on my Mac being connected to wifi.
And it's never finished, in the best way — I'm the client on this one, with my wife, and we keep adding features.