Drafting multilingual replies to customer reviews offline with a small local model

Replying to public reviews matters for reputation but takes time, especially across several languages, and paid reply tools charge a monthly fee per property. Sending customer reviews to a cloud API also raises privacy questions.

1. Run a local llama.cpp server with Gemma 3 4B (quantized) on a normal CPU; no GPU or cloud service. 2. For each review, the first model call detects the language (temperature 0, returns a two-letter code). 3. The second call drafts the reply with a system prompt that works like a reputation-manager playbook: reply only in the detected language, 40 to 80 words, mention one specific detail, apologise without arguing, name a concrete fix for maintenance problems, avoid certain topics, sign as the business. 4. The detected language is restated in the user message. The author's first version just said to reply in the review's language, and the small model answered about half the English reviews in French; detecting first and stating the language explicitly fixed it. 5. Reviews can be processed in bulk from a CSV, or pasted into a small web page for non-technical staff, which returns a draft in about 20 seconds. 6. A person reads each draft (about 30 seconds) before posting. The brand voice lives in an editable prompt.

On 10 real, unanswered Google reviews in several languages, the author checked all drafts: 10 of 10 in the correct language, 10 of 10 specific to the review, none arguing with the guest. He judged them good enough to post after a quick human read. Marginal cost per reply is zero beyond hardware already in place. No posted replies, response rates or rating changes are reported.

Very small test (10 reviews), graded by the author himself. The business is a physical hotel, not an online business; the method should transfer to app store, marketplace or software reviews, but that was not tested. It is a hackathon challenge entry, which may favour a positive write-up. No evidence yet that the hotel or anyone else uses it day to day, or of any effect on ratings or bookings. Quality from a 4B model may vary on harder or hostile reviews.

A computer that can run llama.cpp, the Gemma 3 4B model weights, Python, and comfort running a script. The code is public.

With small models, make hard constraints mechanical: detect the language in a separate call and state it explicitly in the drafting prompt.

• DEV Community (jeffreyturov) ↗ Firsthand write-up, published October 13, 2026, main source
• Codeberg (review-concierge repository) ↗ Code repository, publication date unknown

Source published: October 13, 2026. Last reviewed here: November 3, 2026. Spot a mistake? Tell us.