Sense Signal AI.

The problem

Fragrance is uniquely hard to sell online because shoppers cannot smell before they buy — they rely on notes, descriptions, and reviews that are subjective and easy to misread. Meanwhile, the feedback that would fix this is scattered across reviews, return reasons, support tickets, live chat notes, and surveys, and teams usually spot problems only after customers are already disappointed.

For a retailer like the fictional Luma Scents, the core failure is speed of translation: useful customer feedback exists, but it isn't turned quickly enough into clear product-page content, emotional scent descriptions, or the different insight outputs each internal team needs.

The solution

Sense Signal AI (ScentSignal AI) is an internal feedback-intelligence workflow that finds where a product's promise diverges from customers' actual experience. It groups repeated customer language, compares it against the product-page promise, detects the expectation gap, assigns a confidence/signal-strength score, and prepares a source-evidence summary.

It then runs signal triage — routing each issue to product-page guidance, a Product/Research note, a CX support note, monitor-only, or an investigation flag — and drafts human-reviewable outputs including an emotional Scent Card, expectation and sample-first guidance, and team notes. The governing principle is explicit: AI drafts, humans decide. No customer-facing guidance goes live without approval.

How it works

The MVP uses a single live server-side OpenAI call (GPT-4.1 at low temperature) that returns structured output for the expectation gap, evidence summary, signal classification, routes, draft recommendations, a guardrail check, and a human-review requirement. The API key is stored as a Lovable secret, never exposed in the browser, with a labeled demo fallback if the call fails.

A master prompt enforces evidence grounding (no invented quotes, counts, or scent notes), confidence calibration on weak or mixed evidence, and refusal of unsupported medical, allergy, safety, longevity, or formulation claims. Evaluation on a six-case Signal Triage set started at 67% on a broad grader, then split into three diagnostic graders — Primary Routing Accuracy, Secondary Route Completeness, and Evidence Grounding & Hallucination Safety — each reaching 6/6 after prompt refinement.

Who it's for

Luma Scents is a B2C retailer, but Sense Signal AI's primary user is internal: the CX Insights / Customer Experience Manager who turns product notes and scattered feedback into clear outputs for different teams. Downstream users are shoppers who see approved content on the product page.

Each team in the workflow has clear ownership: CX Insights owns evidence review and recommendation prep; Content/E-commerce approves product-page guidance; CX Lead approves support notes; Product/Research owns product-learning signals; QA owns quality and performance investigations; and Brand/Creative owns Scent Card copy and visuals.

Why it matters

The category is large and digitally influenced: U.S. prestige beauty reached $33.9B in 2024, with fragrance the fastest-growing prestige category (up 12%) and 28% of prestige sales, while 86% of consumers won't buy online without reading reviews. Faster, safer translation of feedback into expectation-setting content directly reduces blind-buy disappointment and expectation-mismatch returns.

Rollout is a controlled internal pilot — historical feedback first, then shadow mode on live feedback, then staged expansion — with all outputs draft-only and RAG introduced only once real product and feedback sources are connected. The scalable version is a monitored, human-approved intelligence workflow, never a fully autonomous publishing system.

At a glance

Project

Sense Signal AI

Built by

Glen Marshall

One-liner

Sense Signal AI is a feedback intelligence workflow for fragrance e-commerce teams that need to understand where product expectations diverge from actual customer experience.