Spotting Trends

A courier changes depots, and late delivery complaints start to climb. Nobody notices for two months, because each support agent only sees their own tickets. If you've labelled your messages, then the whole company can see that shift, in the first week.

• A courier changes depots
• Complaints start climbing
• See it in the first week

Topic counts over time

We count weekly tickets for each topic. Here, in week eight, late delivery complaints jump. We compare the latest week with the first eight, using standard deviations, so that we know how unusual it is. Run it. Late delivery is about six and a half standard deviations above normal. That's a real alarm. Torn cover is flat, at minus point two. Connect your classifier to a simple check like this one, and customer complaints become an early warning system for the entire business.

import numpy as np
import pandas as pd

rng = np.random.default_rng(0)
weeks = pd.date_range("2026-01-05", periods=12, freq="W")
late = rng.poisson(30, 12); late[8:] += 45
torn = rng.poisson(20, 12)
t = pd.DataFrame({"late delivery": late, "torn cover": torn}, index=weeks.strftime("%m-%d"))

base, sd = t.iloc[:8].mean(), t.iloc[:8].std()
print(t.tail(4).to_string())
print(((t.iloc[-1] - base) / sd).round(1).to_string())

       late delivery  torn cover
03-08             78          24
03-15             77          26
03-22             78          18
03-29             78          19
late delivery    6.6
torn cover      -0.2

From alarm to answer

A number says that something changed. Now you need to know why. Pull a sample of the tickets from that spike, and send them to a language model, asking what changed. Group by region and product, to find where. And close the loop, by telling the courier team. A dashboard that nobody acts on isn't worth building.

Summary