RetailMind
GNN-based product recommendation system with MLOps pipeline
Personalization at small-retail scale
Recommendation engines exist for Amazon and Flipkart — they have dedicated ML teams and petabytes of data. Small retailers don't. RetailMind was built to close that gap using the UCI Online Retail dataset (500K+ transactions) as a proxy for a mid-size Indian retailer's two-year purchase history. The harder problem wasn't the model — it was making the pipeline reproducible and production-deployable without a dedicated data team.
Context
- No off-the-shelf recommendation engine for SME retailers
- 500K+ transaction records needing preprocessing
- Research-to-production gap — most GNN work stays in notebooks
- Models must update as new purchase data arrives
Our approach
Key decisions
LightGCN over collaborative filtering
Graph-based model captures second-order purchase relationships that matrix factorization misses — customers who bought A and B together inform recommendations for C.
DVC for full data versioning
Dataset, preprocessing steps, and model artifacts versioned alongside code. Any experiment is reproducible from a single Git commit, not a pinned S3 link.
MLflow for experiment tracking
Every training run logs hyperparameters, NDCG@10, and Recall@20 automatically. Best model is auto-promoted to the production registry — no manual comparison.
FastAPI serving with Swagger
/recommend/{user_id} returns top-N products with relevance scores. Swagger docs generated automatically. Any developer can integrate in under an hour.
Results
What we achieved
NDCG@10 on UCI dataset
Transactions processed
Inference per request
MLOps pipeline with DVC + MLflow
Stack used
Next project
MissionCart →Start a project
Let's build something great
Drop your details below. We'll have it open in Gmail, ready to send — no copy-pasting.