Results at a glance
- Scale: 5,000+ product records monitored and forecast every day — previously a weekly spreadsheet review of a fraction of the catalog
- Cost: ~$0.0005 per product forecast (GPT-4o-mini) — the entire daily run costs less than a coffee
- Control: supplier orders are drafted by the agent, approved by a person — every decision traced and auditable
- Stack: LangChain tool-use agent · GPT-4o-mini · Supabase pgvector · LangSmith
The problem
Inventory in fashion is unforgiving: order too much and cash is trapped in unsold stock; order too little and you miss the season. Pellemoda was managing this the way most SMBs do — spreadsheets, manual reorder points, and experienced staff making judgment calls under time pressure. The knowledge lived in people's heads, the data lived in disconnected places, and a human review could only ever cover a slice of the catalog at a time.
They didn't need a dashboard to look at the problem. They needed something that would act on it.
How it works
The system is a LangChain multi-step, tool-use agent — not a single prompt, but an agent that reasons over the current state, decides what to check next, and acts through real tools. Each daily run:
- Ingests current stock levels and sales history from the company's systems
- Forecasts demand per SKU — historical patterns and seasonality, projected forward, at roughly 1,500 input / 400 output tokens per product (~$0.0005 each on GPT-4o-mini)
- Classifies stock health — flags items drifting toward stockout or overstock, with the why attached, not just the number
- Drafts supplier actions — when a reorder is warranted, the agent prepares the supplier-facing order with its reasoning
- Waits for a human — consequential actions sit in an approval queue; the team reviews the draft and its rationale, then approves in one click
Product data is embedded in Supabase pgvector (~5K records — no dedicated vector database needed at this scale), and every run is traced in LangSmith, so any decision can be audited after the fact: what the agent saw, what it reasoned, what it proposed, who approved it.
Because it's a tool-use agent rather than a hard-coded script, it handles the messy, multi-step judgment calls a rules engine can't — and because the human stays in the loop on orders, the team trusts what ships.
The results
- Coverage went from partial and weekly to total and daily — every SKU, every day, instead of a spreadsheet review of the most urgent items when someone had time.
- Stock decisions start from a forecast, not a reaction — the agent surfaces items drifting off course days earlier than the manual process caught them.
- The institutional knowledge that lived in staff heads — which items spike seasonally, which suppliers need lead time — is encoded in a system that runs every day without being asked.
- Running cost is negligible: the full daily forecasting pass across the catalog costs well under a dollar.
Want the same?
If you have a repetitive, judgment-heavy process that a person does today — forecasting, monitoring, chasing, reconciling — there's a good chance an agent can own it, with your team keeping approval over what matters. Book a 30-minute call and let's find out.