Skip to content
enfa
NLP

NLP Comment Analysis

A pipeline that turns competitors' public reviews into a weekly read on where their product is failing — and where ours was about to.

ROLESenior Product Manager
PERIOD2021 — 2022
CONTEXTE-g Tech Laboratory

THE PROBLEM

Competitive insight arrived as anecdote in quarterly reviews. Meanwhile thousands of reviews were published publicly every week, describing exactly which parts of every product in the category were breaking.

1 quarterthe lag between a competitor's problem appearing and us hearing about it

CONTEXT

Product strategy for a national platform was set on internal telemetry alone, which by definition only sees the users who stayed.

APPROACH

Ingest public reviews across the category, classify by feature area, score sentiment, and report movement week over week rather than absolute levels.

MY ROLE

I defined the feature taxonomy, insisted on trend-over-level reporting, and owned the rule that the system flags questions for humans rather than issuing conclusions.

HOW I WORKED

01TaxonomyEighteen feature areas defined with the product team before any model touched the data.
02PipelineWeekly ingest into BigQuery; spaCy classification; sentiment scored per feature area.
03Trend reportingWeek-over-week deltas surfaced in Looker; absolute scores deliberately de-emphasised.
04Review ritualA thirty-minute weekly read with product leads — the output is a question list, not a verdict.

DECISION LOG

The calls I owned, the alternative I rejected, and the cost I accepted for each.

01
DECISION

Report movement, not absolute sentiment

ALTERNATIVE REJECTEDPublish a competitive sentiment leaderboard
WHYAbsolute sentiment across platforms is not comparable — review cultures differ. Movement within one source is a real signal.
COST ACCEPTEDNo headline ranking, which is what leadership initially asked for.
02
DECISION

Human-defined taxonomy before modelling

ALTERNATIVE REJECTEDUnsupervised topic clustering
WHYClusters that don't map to owned feature areas produce insight nobody can action. The taxonomy made every finding assignable.
COST ACCEPTEDThree weeks up front and a blind spot for genuinely novel themes.
03
DECISION

Output questions, not conclusions

ALTERNATIVE REJECTEDAuto-generate recommendations
WHYThe pipeline sees complaints, not causes. Framing output as questions kept the interpretation with the people who had context.
COST ACCEPTEDLower perceived sophistication in demos.

WHAT MOVED

BEFOREAFTER
Insight latency1 quarter1 week
Feature areas monitored018
Sources ingested weekly—6

OUTCOME

18feature areas tracked week over week
1 weekfrom public complaint to product review
6competitor sources in the pipeline

WHAT I PRODUCED

Feature-area taxonomyIngest & classification pipelineWeekly trend dashboardReview ritual format

WHAT I'D DO DIFFERENTLY

I built the dashboard before the ritual. Adoption only started once the weekly thirty-minute read existed — the meeting was the product, the dashboard was the artefact.

Related writing

Routing 500 daily deliveries without a routing teamThe constraint model behind next-day delivery at national scale.Instrumenting a product for 10M merchantsEvent design that survives its first million users.
NEXT PROJECTAI Customer ServiceDeployed LLM-powered system for Bank Tejarat affiliate, handling enterprise client queries and categorizing tickets automatically.