case 03 / 05

Lumina Vision CRM

Real-time facial analysis and customer sentiment tracking for physical retail.

Year

2024

Role

Serverless · Realtime UI

Status

Deployed on private retail networks

( architecture )

How it's built

Kinesis VideoAWS RekognitionLambdaDynamoDBAppSyncReact

Hover a service to trace its connections

ClientEdge / APIComputeEvents / queuesStorageAI / ML

( the problem )

Store managers had sales numbers but no idea how customers felt while they were in the aisle, and data arrived a day late.

( the build )

Camera streams are sampled at the edge, analysed by Rekognition and pushed as anonymised sentiment events to a live React dashboard over AppSync subscriptions.

300ms

camera → dashboard

stores streaming live

100%

anonymised, no faces stored

( request flow )

01In-store cameras publish to Kinesis Video Streams.
02A sampler Lambda extracts frames and sends them to Rekognition.
03Emotions are aggregated per zone and written to DynamoDB, never the image.
04AppSync subscriptions stream every update to the React dashboard.

lumina.console● live

300ms

camera → dashboard

stores streaming live

100%

anonymised, no faces stored

handler.ts

1const faces = await rekognition.detectFaces({

2 Image: { Bytes: frame },

3 Attributes: ['EMOTIONS'],

4});

5publish(zone, aggregate(faces));

6▍

main ✓TypeScript · UTF-8 · Ln 6

Next case · 04

Velocity Oracle →