From 30 minutes of manual handling to a few minutes - built with Make · Gemini · AppSheet · Google Data Studio
A real-estate developer managing 18 concurrent projects handled 800 tenant requests a month through a fully manual process - emails, WhatsApp messages and letters passed between field coordinators with no synchronization. I built an end-to-end system that identifies the tenant, classifies request urgency using AI, and automatically routes it to the right handler - with safety nets for every case the AI isn't confident about.
Unstructured information (free text, images) combined with a subjective decision (severity level) - a combination that resists traditional hard-coded rules and requires natural-language understanding.
Incoming email → tenant & team identification (deterministic logic) → extraction & classification (Gemini Flash) → routing to 4 paths (critical / urgent-regular / manual review / technical failure) → write to Sheets + matching alert. An AppSheet layer on top gives the field team a day-to-day working interface from their phone.
Chose Resume over Rollback for API-failure handling - for a non-critical request, halting the entire flow costs more than routing it immediately to manual review.
Built the AI decision-reliability measurement to be bidirectional, not one-directional - so a case where the AI was too cautious is also caught, not only a case where it was too lenient.
Split one dataset into two separate KPIs (historical vs. operational Backoffice rate) - because "how often has AI ever been wrong" and "how much work is queued right now" are two different questions needing two different answers.
Distinguished AI_urgency from Urgency in the accuracy calculation - one frozen field, one that updates - so the success metric doesn't drift as more tickets get closed.
Backoffice rate: 23.7% historical cumulative, 9 tickets currently queued
Classification Delta: 6.9%-20% weekly, vs. an 85% accuracy target for production go-live
Time to confirmation: down from 30 manual minutes to a few minutes in the pilot
