All case studiesRetail

AI Product Recommendation Engine for D2C Fashion Brand

Built a collaborative-filtering + LLM hybrid recommendation system. Average order value increased 38% in 3 months post-launch.

⏱ 12 weeks👥 3 ML engineers + 2 backend engineers

Results

+38% average order value within 3 months

+22% click-through rate on recommendation widgets

Sub-80ms p99 inference latency

A/B test significant at 95% confidence

Cold-start coverage: 100% of new users receive relevant results from session 1

The challenge

A D2C fashion brand with 180,000 monthly active users was using a basic "frequently bought together" widget that drove minimal incremental revenue. Their catalogue of 12,000 SKUs changed seasonally, making static rule-based approaches quickly stale. They needed a recommendation engine that could handle cold-start users, respect inventory constraints, and explain its suggestions in natural language.

Our approach

1
Audited existing data: purchase history, browse events, wishlist actions, and return data.
2
Built a two-stage retrieval system: collaborative-filtering (ALS) for candidate generation, LLM re-ranker for final ordering with style-coherence reasoning.
3
Integrated real-time inventory signals via Kafka to suppress out-of-stock or low-stock items.
4
A/B tested using a 50/50 traffic split over 6 weeks with Optimizely.
5
Exposed recommendations via a low-latency FastAPI endpoint (p99 < 80ms) consumed by the Next.js storefront.

Tech stack:PyTorchNext.jsMLflowKafkaFastAPIOptimizely

Want a similar outcome?

Tell us about your challenge and we'll scope a solution within 48 hours.

Start a conversation