My Parents Ran a Noodle Restaurant for 15 Years. They Never Read a Single Google Review.

Built for the Hacktoberfest 2026 Weekend Challenge — "Build for a Friend". Entered categories: Best Use of Gemma, Best Use of SerpApi.

My parents run a small noodle restaurant in the San Gabriel Valley. They have run it for fifteen years. In those fifteen years, they have never read a single one of their own Google Maps reviews — because every review is in English, and they don't read English.

The 1-star reviews sit there, unanswered. Every restaurateur knows that replying to reviews is what turns a bad rating into returning customers. But you can't reply to what you can't read.

So I built them a review secretary.

What it does

review-bridge watches Google Maps for them and runs a loop they can actually use:

盯着 → 汇总 → 预警 → 代笔回复 — watch, brief, alert, ghostwrite.

1. 📊 中文简报 (brief) — every morning, a Chinese briefing: average rating, rating trend, which dishes customers praise most, what they complain about most. Google Translate can translate a review. It cannot do this.
2. 🚨 差评预警 (alert) — 1–2 star reviews get flagged immediately with urgency. My parents never open Google Maps; the tool watches it for them.
3. ✍️ 代笔回复 (ghostwrite) — for each flagged review, a polite, natural English reply draft they can post publicly, plus a Chinese explanation of what the reply says and why. This is the real unlock: non-English-speaking owners responding to customers in fluent English.

Translation happens underneath, as plumbing. It is never presented as the product.

Why open AI matters here

Everything runs on the family laptop, offline:

• No internet needed in a restaurant backroom — the AI core is Gemma 3 (gemma3:4b) running locally via Ollama.
• Customer feedback never leaves the laptop — no review data sent to a third-party AI API.
• Zero cost — a small family business shouldn't pay a SaaS subscription to read its own reviews.
• The family controls it — the whole web UI is in Simplified Chinese, tuned for the owners, not for a generic dashboard.

Review fetching uses SerpApi's Google Maps Reviews API — the free tier gives 100 searches/month, which is plenty for one restaurant. No SerpApi key? There's a paste mode: paste review text in, get the briefing out, no network needed at all.

The whole thing is dependency-free Python (stdlib only, 26 tests), installable with one pip install review-bridge and a one-click PowerShell setup script for Windows.

The real test

I ran it on the family Windows laptop: Ollama with Gemma 3 (4B) running fully offline, plus a SerpApi key for fetching real Google Maps reviews. The --mock demo worked first, then a real query against our own restaurant — 8 reviews, 4.9 stars — and the Chinese briefing came out clean, even catching the one recurring complaint (dishes arriving lukewarm when the restaurant gets crowded).

Then the stress test: a nearby Burger King with a 2.9-star rating. Three of its eight reviews were 1–2 stars. All three got flagged, each with a Chinese translation, a polite English reply draft, and a Chinese explanation of what the reply says and why.

I showed it to my parents on the family laptop, in Chinese, on their phone's browser.

"哎呀，儿子长大了，真好，会体谅我们了，还能做网页了，真厉害。"

"Oh my, our son has grown up. That's wonderful — he really looks out for us now. And he can even build web pages. So impressive!"

— my mom, seeing the Chinese briefing of our own restaurant's reviews for the first time in fifteen years

Real briefing: 4.9 stars across 8 reviews, generated locally by Gemma 3.

Bad-review alerts from the Burger King stress test: each flagged review gets a Chinese translation, an English reply draft ready to paste into Google Maps, and a Chinese explanation of the reply.

The full loop in --mock mode: briefing, alerts, and reply drafts — no API keys, no model, no network.

Prize Categories

• Best Use of Gemma — the entire AI core (briefings, alert triage, reply drafts, explanations) runs on Gemma 3 (4B) locally via Ollama. No cloud inference, no API costs, customer data never leaves the laptop.
• Best Use of SerpApi — real Google Maps reviews are fetched through SerpApi's Google Maps Reviews API (the free tier's 100 searches/month is plenty for one restaurant), grounding every briefing in fresh, real-world data.

Honest limitations

• The SerpApi free tier caps at 100 searches/month; a busy restaurant chain would outgrow it.
• Reply drafts are AI-generated — the owners should skim the Chinese explanation before posting.
• Star ratings on aggregator mirrors are sometimes sub-ratings; the tool trusts what the API returns.
• The local model needs ~4GB RAM free; a very old laptop will be slow (the --mock demo still works).
• This watches Google Maps reviews only — Yelp and DoorDash are separate worlds.

Try it

• GitHub: https://github.com/hahahahahahahahah6/review-bridge
• PyPI: pip install review-bridge
• Demo: review-bridge serve --mock — no keys, no model, no network

If your family runs a small business and the internet talks past them — build the bridge. It's a weekend of work, and fifteen years of unanswered reviews is a long time.