Signal From Noise: How AI Reads a Class WhatsApp

AI can write code and summarize books, yet parents still scroll 200 messages to learn tomorrow's pickup moved to 2:30. Why? Because a group chat is an unstructured, multi-speaker, mixed-intent text stream — one of the messiest inputs in NLP. Here's the pipeline that turns chat chaos into a calendar entry.

The message stream — drag to rotate2 / 12 flagged

Extracted structured facts

The three-stage pipeline

1

Intent classification

Each message gets scored per category: logistics, social, lost & found, question. "Happy birthday Maya!!" scores ~0.98 social / 0.01 logistics. "Heads up, early dismissal Thu" scores the reverse. Modern LLMs do this zero-shot; older systems needed thousands of labeled examples.

2

Entity & slot extraction

From flagged messages, the model pulls typed fields: WHAT (dismissal), WHEN ("Thu" → resolved to an actual date using the message timestamp), WHO (Year 4), ACTION (adjust pickup). Ambiguity is the hard part: "next Friday" means different dates to different people.

3

Conflict resolution

Group chats self-correct: "Actually scratch that, trip is the 14th not the 12th." The system must track that message 87 supersedes message 62 — a coreference and temporal-ordering problem. Naive keyword alerts fail exactly here.

Why this is genuinely hard: a 30-family class chat produces roughly 150–300 messages a week; studies of group-chat corpora find under 10% carry actionable logistics. That is a needle-in-haystack ratio where both error types hurt — a false negative means a kid stranded at pickup, and false positives train parents to ignore the assistant. The precision/recall threshold slider above is the real product decision every "family logistics AI" has to make.

Enjoy this tool? Build your own with Super