🍼 Baby Feed

A home-screen widget for new parents who are too sleep-deprived to open an app and do math.

Built in a weekend for Hacktoberfest 2026 — "Build for a Friend" (theme: open-source AI at its core), as a gift for a close friend whose newborn arrived a few weeks ago.

My friend was tracking every feed in an Obsidian note. Figuring out "when was she last fed, and when should I feed her next" meant unlocking his phone, opening the app, finding the last entry, and doing the math by hand — real friction at 3am with a newborn.

He needed to log a feed and see the next feeding time from his home screen, without opening anything.

• One-tap logging. The widget shows last feed time, next feed time, and a live countdown. Tapping "Log feed now" records the current time — no app, no typing.
• Overdue, not broken. Once the countdown passes zero, the widget switches to an honest "Overdue by 18m" instead of showing a stale or negative number.
• Voice/text backdating, for the one time you forget. Open the app, say or type "she fed about 20 minutes ago" or "fed at 1:30", and an on-device model figures out the timestamp. You confirm before it saves.
• History & settings. Edit or delete a mistaken entry, and adjust the feeding interval as the baby's schedule changes with age.

This was built for Hacktoberfest's "open-source AI at its core" theme, and the AI genuinely had to earn its place — not be bolted on for the submission.

Backdating a feed needs to understand phrases like "about 20 minutes ago" or "fed at 1:30" and turn them into an exact timestamp. That's a natural-language problem a simple form can't solve gracefully, and it's exactly where a small open-weight model is a good fit — not a chatbot, not a cloud API, just a narrow extraction task running locally.

• Everything stays on the phone. Both speech-to-text and time parsing run entirely on-device — no network call is ever made to log a feed. Nothing about a newborn's feeding schedule leaves a device my friend doesn't fully control.
• Works at 3am with no signal. No cloud dependency means no "service unavailable" moment during a feed in the middle of the night.
• Costs nothing to run. No API key, no per-request bill, no usage quota — the model runs for free, forever, on hardware he already owns.
• The model does one honest job. Rather than asking the LLM to do date arithmetic (which, at 1B parameters, it turned out to be unreliable at — see below), it only classifies the input ("RELATIVE 20 MINUTES", "ABSOLUTE 1 30 PM"); Kotlin does the actual subtraction deterministically. The open model is doing what open, inspectable, swappable software is good at: a small, well-scoped job you can reason about and fix when it's wrong.

The first version asked the model to compute the final timestamp directly (e.g., resolve "20 minutes ago" against the current time itself). On real hardware, Gemma 3 1B reliably understood the phrase but consistently got the subtraction wrong, defaulting to "just now" regardless of few-shot examples or decoding temperature. The fix wasn't a bigger model — it was recognizing that arithmetic belongs in code, not in a language model's forward pass. The model now only extracts structure; java.time does the math. It's a small example of why "open" matters here beyond ideology: when something didn't work, the fix was visible and fixable, not a black box to work around.

Backdate a feed by voice or text	History — edit or delete entries	Settings — adjust the feeding interval

The app is intentionally narrow in scope: feeding only (no diapers/medicine/massage tracking), Android only, single device, no cloud sync, no notifications. See the original requirements doc for the full reasoning behind every one of those cuts.

Requirements: a recent Android Studio, a physical Android device running API 31+ (the on-device SpeechRecognizer APIs this app relies on don't exist on the emulator), and a Hugging Face or Kaggle account.

1. Clone and open the project

   git clone https://github.com/nish17/baby-feed.git
   cd baby-feed

   Open it in Android Studio and let it sync.

2. Download the model (one-time, ~555MB)

   Gemma models require a free license click-through before download. Grab gemma3-1b-it-int4.task from litert-community/Gemma3-1B-IT on Hugging Face (accept the Gemma Terms of Use first).

3. Push the model to your device

   adb shell mkdir -p /data/local/tmp/llm
   adb push gemma3-1b-it-int4.task /data/local/tmp/llm/model.task

   The app looks for the model at exactly that path. If it's missing, backdating silently falls back to a manual time picker instead of crashing — the app is fully usable without ever doing this step, just without AI-assisted backdating.

4. Build and install

   ./gradlew installDebug

   Then long-press your home screen → Widgets → Baby Feed to add the widget.

• 53 unit tests — business logic, view models, and the Glance widget composable, run entirely on the JVM (no device needed): ./gradlew testDebugUnitTest
• 14 instrumented tests — Room DAO, WorkManager workers, on-device speech recognition, and real on-device LLM inference against the pushed Gemma model: ./gradlew connectedDebugAndroidTest

Built for Hacktoberfest 2026's "Build for a Friend" weekend. Thanks to the open-weight model and tooling that made the on-device AI pitch possible: Gemma 3 and MediaPipe from Google.

And mostly — thanks to the new parents this was actually built for. Congratulations. 🎉