Multi-agent workflow automation: 30x efficiency and 5x growth to $12M ARR

Replaced a hand-curated merchandising process with a multi-agent workflow built on hybrid retrieval, re-ranking and a fine-tuned model — with merchandisers in the loop as the eval signal.

Across 100+ retailers, the rate-limiting step on growth was a small team of merchandisers hand-building assortments. Quality was inconsistent, and the best account, a key client, was capped at what humans could curate.

The merchandiser override isn't an inconvenience; it's the highest-signal eval we have. Treat it as the production quality metric and the system improves itself.

1. 01

   Built an agentic workflow that turned a merchandising brief into a finished assortment, grounding generation in the retailer's product catalog, seasonal trends, and brand style guides, in place of a manual team of five.

2. 02

   Shipped fast by keeping merchandisers in the loop: every AI-generated assortment went through human review first, which protected brand quality and built client trust, then layered in an AI judge-and-fixer that caught discrepancies and progressively removed the manual step.

3. 03

   Made two mid-flight engineering pivots to hold the quality bar at scale: moved from keyword-only RAG to hybrid retrieval (keyword plus embeddings, with re-ranking) when precision fell short, and transitioned from a large model to a fine-tuned smaller one to cut cost.

4. 04

   Built the eval system that made it trustworthy: scored each stage (concept extraction, assortment quality via attribute-level Jaccard against a gold set, and hard-filter enforcement), with the merchandiser override rate as the live production quality signal.

A multi-agent workflow replaced manual curation and ran 30x more efficiently, saving $2.3M a year. It scaled a key account 5x, from $2.5M to $12M in ARR, with zero quality escalations. The same infrastructure was later reused to overhaul the consumer discovery experience with semantic search and personalization, driving a 300% conversion lift.

"The pattern I now reach for first with AI: ship it with a human in the loop to earn quality and trust, then optimise for scale and progressively automate the review away. It became our template for every AI build that followed."