Manyavar, one of India’s larger clothing chains, had rich data on the back-end of their sales — bills, inventory movement — but nothing on the front-end. They could tell you what sold, not who looked at what before deciding, who walked past that rack three times, or whether people were coming in solo or in couples or in families. All the retail decisions — what to merchandise, how to lay out a store, how to staff — were being made with half the information missing. I built Argust as the other half.
The pipeline
Argust reads from the store’s existing CCTV feeds — no new camera hardware — and pipes them to a local GPU server, where the vision work runs on-prem rather than in the cloud (both for latency and because retail footage doesn’t belong off-site). The pipeline does a handful of things in parallel:
- Identity assignment — every visitor gets an anonymous numeric ID on first entry
- Re-identification across days — the same shopper walking in next week is tagged as a revisit
- Demographics — rough age bucket, presented gender, ethnicity signals where relevant to merchandising
- Group detection — are they alone, a couple, a family of four, a group of friends? This was the thing Manyavar most wanted
- Journey tracking — which sections did they move through, how long in each, what order
The correlation layer
The interesting output isn’t any single metric — it’s the join with the sales data. If a shopper browsed section A, then B, then bought from C, that’s a data point about what the store’s layout is actually doing. Do families buy differently than couples? Does dwell time in a section predict purchase, or is it inverse? Merchandising decisions that used to be intuition became measurable.
Status
Built and demonstrated as a POC for Manyavar. They ended up awarding the production contract to a competitor who’d pitched earlier. Argust didn’t ship commercially, but the pipeline was fully working.