Built a crisis-alert system that outpaced a 150-person monitoring team危機預警系統，偵測速度超越150人監控團隊

Situation

A global e-commerce company needed earlier warning of emerging events that could affect its operations or reputation. The trigger was a period, during the COVID-19 pandemic, when a sharp rise in price-gouging on essential health and hygiene products on the platform drew significant regulatory and media scrutiny. Detection at the time relied on a manual monitoring team of roughly 150 people tracking online mentions, so by the time an issue became visible, exposure had often already grown considerably.

The initial ask was straightforward — monitor brand-related conversation online and alert PR and strategy teams when negative volume rose — but volume and average sentiment alone made poor early-warning signals: a global brand generates heavy discussion every day, most of it ordinary background noise, while genuinely risky events could start small and accelerate quickly around a narrow topic or platform.

Strategy

I scoped and led the design of a crisis detection and alerting system that combined NLP classification with time-series forecasting. Working with the client’s brand and communications teams, I translated a broad monitoring request into three measurable capabilities: identifying potential crisis events, forecasting their trajectories, and converting those forecasts into signals the business could act on.

1. Identifying potential crisis events using NLP.

The first challenge was limited training data. Crisis events vary considerably by topic, sentiment, product category, platform, and external context. A small static sample could not adequately represent the range of scenarios the classifier would encounter in production.

Rather than training a crisis classifier directly on this limited sample, I designed a two-stage approach. We first trained a broader negative-sentiment detector and used it to identify a substantially larger pool of potentially relevant content. That expanded dataset was then reviewed and labelled to train the more targeted crisis-event classifier.

This approach increased the diversity of the training data and reduced the risk of producing a model that performed well on a narrow offline test set but failed to recognise previously unseen crisis patterns in production.

2. Forecasting how events were likely to develop.

Identifying negative content was not sufficient. Static sentiment scores and raw mention counts generated false alarms for high-volume topics while missing smaller issues that were beginning to accelerate.

I therefore designed a time-series forecasting layer that used each topic’s recent activity to project its likely trajectory. The system flagged topics whose forecast activity was accelerating sharply relative to their own historical baseline.

This distinction was central to the solution: the system assessed where a topic was heading, rather than reacting only after negative conversation had already reached a fixed threshold.

3. Translating model outputs into operational signals.

Raw model probabilities would have required business users to interpret the analytics themselves. I converted the classification and forecasting outputs into a set crisis-risk metrics and topic category that PR, brand, and CRM teams could use directly.

The resulting alerts gave teams a common basis for deciding whether to escalate an issue, adjust a live campaign, prepare a response, or continue monitoring—without requiring them to understand the underlying models.

Result

The system identified critical events approximately 3–6 hours earlier than the existing 150-person manual monitoring process, giving PR and brand teams a meaningful window to respond before an issue reached wider visibility.

Its outputs were incorporated into CRM and campaign-response decisions, enabling the client to adjust messaging or pause campaigns associated with an emerging issue instead of reacting after the event had already escalated.