Signal Engine

Support teams are the first to know when a product breaks — and the last to be heard. The evidence lives in thousands of unstructured conversations, so when it finally reaches a roadmap discussion it arrives as anecdote: “customers are mad about billing.” Product hears an opinion. Engineering checks the tracker and sees “low severity.” Nothing moves.

The tooling makes this worse, not better. Manual tags systematically undercount reality — in one audit, an issue a team believed was tagged appeared in just one of six relevant tickets. Dashboards answer “what happened last month,” never “is right now normal?” And surveys hear from almost nobody: typical B2B response rates sit near 12%, while every one of those silent customers is already explaining exactly what’s wrong — in the queue.

We had also lived the manual alternative. In a department handling 3.5 million tickets a year, building one reliable insight pipeline took a team of ten roughly six months and $50,000 in labor — three-week cycles, seven to ten iterations to reach target coverage — and it started going stale the day it shipped.

Signal Engine ingests every support conversation as it lands and reads each one the way a good analyst would — then never gets tired and never skims. It grows a hierarchical category tree out of the raw text, scores every conversation against a structured rubric, and layers statistics on top so that what reaches a human is signal, not vibes.

The work surfaces in three ways. A daily Insight Feed briefs the team each morning like an analyst who worked overnight — alerts grouped by severity, each carrying the “why,” the blast radius, and a recommended action. Drill-down analytics connect any alert to the exact conversations, customer quotes, and impact metrics behind it. And an agentic chat interface answers ad-hoc questions, runs multi-step investigations, and can be @-mentioned like a teammate.

The architecture came out of a hard lesson learned building chat systems at scale: paste 500 tickets into a chat window and ask for trends, and you get an answer that sounds reasonable and is mathematically wrong. Language models are great at language and bad at math — ask one for the average sentiment of a thousand interactions and it will guess “7 out of 10,” because that’s the probable answer, not the computed one.

So Signal Engine uses the model only where fuzzy interpretation genuinely helps, and plain code everywhere else. Pointed at messy human language, the property people fear as hallucination is the useful one — imagination, the interpretive leap that turns “I keep clicking the button and I’m losing my mind” into a typed data point: high user effort.

On top of that store, the detector separates fires — a metric far outside normal today — from slow burns, the four-day drifts that become next week’s crisis, and ranks everything by severity, reach, and duration. A quietly worsening week outranks a one-day spike. It surfaces positive anomalies too, because a team should know when the numbers reflect its good work.

Just as deliberately, the system says what the data cannot answer — naming the missing dataset instead of manufacturing confidence. An instrument that answers every question isn’t a measurement instrument at all.