Own product / WhatsApp + AI
The customer sends a message with a photo, the AI classifies the fault, a person approves the price and a verified local tradesperson is found. Nothing to download and no ringing round five companies.
Project sheet
unrato.es
Finding a plumber for a dripping tap is harder than it should be: calls nobody answers, quotes that never arrive and prices you only learn at the end. For the self-employed tradesperson, the problem is the reverse: jobs come in with no photo, no address and no idea whether the customer will accept the price.
02 / Solution
Un Rato runs on the official WhatsApp API, n8n and Postgres. The AI has a narrow role: it writes the conversation, extracts the data and classifies the fault from up to three photos. The decisions that matter are not its to make. Status changes are decided by deterministic code, the price comes from a catalogue and a person approves it from Telegram before it reaches the customer.
Webhook with verified HMAC signature and deduplication by message ID
Per-phone lock in Postgres so a text and a photo sent together do not open two requests
A six-second wait to group messages before replying
Vision-based classification with JSON output against a strict schema
A Telegram card with approve, change price, visit or reject
The model never sets prices or changes a request's status. It returns a service code and the system looks up the price in the catalogue.
Every price is approved before it is sent, and the operator can take over the conversation at any time and hand it back to the bot afterwards.
Tradesperson onboarding is a state machine with no AI. The profile starts inactive and the database blocks activation until a person has verified the self-employed registration and public liability insurance.
03 / Scope
04 / Impact
Today, WhatsApp intake, photo classification, human price review and tradesperson onboarding are all live. Automatic assignment to a tradesperson is not yet: the offer and confirmation templates are under review at Meta and, in the meantime, the operator gets an alert to find one. The system already measures what will be needed to decide whether to scale: time to first response, price held and bot accuracy against the operator.
05 / Questions
The price does not reach the customer until a person approves it. If the model has low confidence, the system proposes a diagnostic visit instead of a price, and if the fault is not in the catalogue, the conversation goes to the operator.
Because the review has to happen in seconds from a phone, with the photo in front of you and four buttons. Building a web dashboard for that would have delayed launch without improving the decision.