An interactive pour-over brew timer that runs as a web app / phone PWA. Pick a method (Balanced, Hoffmann, Kasuya, or a custom recipe), set your coffee dose, and it walks you through the pours with a timer, ring, and prompts. It also has an AI Barista chat for troubleshooting your cup ("too sour", "draining too fast", etc.).
The whole app is a single index.html file, hosted free on GitHub Pages.
- Install the V60 integration firmware from the CoffeeScale repo on the scale, and calibrate it with a known weight. Put your Android phone and scale on the same WiFi network.
- For the initial pairing, choose WiFi scale, tap Find scale, and select CoffeeScale from Android Chrome's Bluetooth picker. The app reads the scale's WiFi address automatically, disconnects Bluetooth, and connects over WiFi. Allow local network access when Chrome asks. No IP copying is needed.
- Select your recipe and dose, then tap Scale brew.
- Tap Tare & start. The first pour starts the scale's timer automatically.
- Track the current phase's added / required grams and elapsed / allotted time. For example, the second Balanced pour at a 15g dose might show 70 / 95g and 0:10 / 0:30. Once its water target is reached, it shows Wait while the same phase clock continues to the next pour's start time. The next phase resets the counters. Cumulative scale weight and total brew time remain visible below. Expand View recipe for the full schedule. Missed targets stay active with overtime shown; drawdown has its own timer. Tap Finish brew after drawdown; Cancel or leaving for Settings also ends the board session. A slow brew continues past the recipe's suggested finish.
After that, turn on the scale and open the app. It remembers the pairing and
automatically reconnects over WiFi. It never opens the Bluetooth picker during
ordinary startup or reconnection. If the scale is still starting, the app waits
and retries. It tries both the last address and the stable coffeescale.local
name; after a DHCP address change, a successful local-name connection updates
the remembered IP. Modern Android supports .local resolution, but the router
must allow local multicast for that recovery path. The scale's persistent device
ID prevents another scale at an old IP from supplying your brew readings.
Address entry is available inside Connection troubleshooting if needed.
The board owns pour detection and elapsed time. The app polls one request at a
time, waits 200ms between successful readings, and shows a connection error instead
of stale weight. After a connection gap it catches up to the scale's timer. Keep
the PWA open during brewing. Use a current Chrome on Android; older browsers may
block an HTTPS page from reaching the HTTP scale. If permission was denied, allow
local network access for https://ocoomber.github.io in Chrome's site settings.
Bluetooth is used only for initial discovery, not live weight or timer control.
Close the scale's other live weight page while using V60 to avoid competing polls.
The firmware permits this GitHub Pages origin; hosting elsewhere requires updating
the origin allowlist in CoffeeScale. Manual timer remains available.
The integration uses GET /api/weight, POST /api/pour/arm (tare and arm) and
POST /api/pour/end. It does not change the scale's saved espresso mode.
App logic checks: node --test tests/scale.test.cjs. These simulate the scale API
and Bluetooth picker. Android pairing, permission handling, physical pour detection and live WiFi reliability
still need testing with the flashed scale. After an app update, reload
https://ocoomber.github.io/v60/ in Chrome and reopen the PWA.
Think of the AI Barista as a little coffee-advice hotline.
- The app on your phone = the phone in your hand. Just buttons and a chat box. It can't think — it shows things and passes your question along.
- Google's Gemini AI = a world-expert barista in a back office. This is the bit that actually knows things and writes the answers. It lives on Google's computers, not your phone. You never talk to it directly.
- The Cloudflare Worker = a trusted receptionist sitting in between. Every question goes to her first; she walks over to the expert, asks, and brings the answer back.
To talk to the expert you need a secret pass (the API key) — like a membership card that bills questions to your account. But the app is public: anyone can read its code, like a menu in a shop window. If the key were in the app, it'd be like writing your PIN on that menu — strangers could copy it and use your membership.
So the key lives with the receptionist, behind a locked door. Your phone never sees it. The receptionist also:
- Only takes calls from your app (ignores strangers ringing direct) — "CORS".
- Clips the house rules to every question: "You're a coffee expert. Only answer coffee questions; politely decline anything else." That's why it says "sorry, just coffee." We keep this rule on her side so nobody can erase it.
- The app page → GitHub Pages
- The receptionist → Cloudflare (
barista-worker/) - The expert → Google Gemini
Your phone → GitHub (the page) → Cloudflare (the receptionist) → Google (the expert), and the answer comes back down the same chain.
The limit is counted in requests per day, not words. Roughly:
| Limit | Number | In plain terms |
|---|---|---|
| Requests per day | ~1,500 | One chat message = one request. So ~1,500 questions a day, shared across everyone using the app. |
| Requests per minute | ~15 | You can't fire more than ~15 in a single minute (you'd never hit this by hand). |
| Tokens per minute | ~1,000,000 | "Tokens" are chunks of words (~¾ of a word each). A short Q&A is a few hundred tokens, so this ceiling is effectively unreachable for chat. |
In practice: for you and a few friends, you'll never get close. If a busy day ever did use up the ~1,500, the Barista just says "come back later" and the count resets the next day. Because there's no card on the account (free tier), hitting the limit costs nothing — it pauses, it never bills.
Short answer: it's a strong lock, not a vault.
The house rule (the system prompt) is what makes the bot refuse non-coffee questions, and modern models follow it well — your "ignore all previous instructions" attempt bounced off, which is the model doing its job. But tricking a chatbot into ignoring its instructions is a known, unsolved cat-and-mouse game called prompt injection / jailbreaking. A determined person with clever enough wording can sometimes get any chatbot to slip. No system prompt is 100% guaranteed.
Why we don't lose sleep over it here: even in the worst case where someone jailbreaks it, the damage is tiny —
- they can't see or steal your API key (it's behind the locked door),
- they can't run up a bill (free tier — no card),
- the only thing they could do is waste some of the daily free question-tickets.
So the blast radius is "a stranger wasted a few of today's free questions," which resets overnight. For a personal brew app that's a fine trade-off. If it ever became a public app with real traffic, we'd add stronger guards (rate limits, a separate "is this coffee?" check, etc.) — but that's a bridge for later.
- The Barista proxy and its one-time setup steps live in
barista-worker/. - The app is just
index.html— open it in a browser, or it auto-updates on GitHub Pages when changes are pushed.