Insights Engine
Reads hundreds of Amazon reviews per SKU and tells FMCG brands which product features are driving satisfaction or complaints.
One-star reviews get weighted five times higher than the rest, since that's where real product risk hides.
The problem
FMCG brands with large product catalogs had no visibility into which products were winning, which were quietly failing, or why. The real answers were buried across hundreds of unread Amazon reviews per SKU. Insights Engine reads those reviews at scale and tells brands which product features are driving satisfaction or complaints, and which lessons from the top performers should carry across the portfolio.
The guardrail
A single LLM pass over thousands of reviews goes vague, and classic topic modeling can't tell you which exact feature (battery, packaging, taste) is driving sentiment. Map-reduce summarization paired with aspect-based sentiment analysis (ABSA) keeps the output feature-specific even at thousands of reviews per brand, and collection stays reliable even against Amazon's active bot-blocking.
Since the output feeds real business decisions, it can't hallucinate: every numeric claim is validated against the source data, ABSA auto-corrects invalid aspect tags, and scoring is penalty-adjusted so a handful of loud one-star reviews can't skew a product's ranking.
Stack
Python · Playwright · BrightData · map-reduce · FastAPI · Supabase · ABSA
I built this after watching a founder scroll through 200 reviews by hand, looking for one pattern. Code's not public, it runs on paying brands' review data.
Later integrated as a real-time analytics dashboard inside TruthSeek, used by product-growth teams at FMCG brands.