Data & ML · Advanced
Two-Tower Recommender Engine
Build a production recommender with two-tower retrieval and a ranking model, serve nearest-neighbor candidates from an ANN index, and run an offline-to-online A/B eval loop.
You build a full two-tower retrieval plus reranking pipeline on a public interaction dataset (MovieLens or Amazon Reviews), then serve it behind a FAISS approximate-nearest-neighbor index with a BentoML HTTP endpoint. The project covers the full MLOps loop: offline training tracked in MLflow, online candidate retrieval, a pointwise ranking model that rescores the top-K candidates, and an A/B evaluation harness that compares ranking strategies using recall, NDCG, and online click-through proxies. It is worth building because two-tower retrieval is the production architecture behind YouTube, Pinterest, and TikTok recommendations, and most tutorials stop before the serving and eval layers.
What you build
- Train separate user and item embedding towers in PyTorch and export them independently
- Build and persist a FAISS IVF-PQ index over all item embeddings for sub-millisecond ANN retrieval
- Implement a pointwise ranking model that rescores the top-K FAISS candidates using dense features
- Track every experiment run (embeddings, hyperparameters, retrieval metrics) in MLflow
- Serve the retrieval and ranking stack as a single BentoML service with a /recommend endpoint
- Run an offline A/B eval loop comparing random, popularity, retrieval-only, and full pipeline strategies on held-out data
- Export quantized ONNX towers so the index can be rebuilt incrementally as new items arrive
What it teaches
- Two-tower architecture design: separate encoding of query and item spaces with in-batch negative sampling
- Approximate nearest neighbor indexing with FAISS: index types (Flat, IVF, PQ), trade-offs between recall and latency
- MLOps experiment tracking with MLflow: logging metrics, artifacts, and model registry promotion
- Reranking as a second-stage model: pointwise scoring over a small candidate set vs. full corpus
- Offline evaluation methodology: train/val/test splits by time, ranking metrics (NDCG, MRR, recall@K)
- Production model serving with BentoML: packaging multi-model pipelines into a single versioned service
How it works
- 1
Raw Interactions
- MovieLens events
- user/item features
↓ preprocess
- 2
Two-Tower Training
- user MLP tower
- item MLP tower
- MLflow tracking
↓ embed
- 3
ANN Index Build
- item embeddings
- FAISS IVF-PQ
↓ retrieve
- 4
Online Retrieval
- user query vector
- top-K candidates
↓ rescore
- 5
Ranking Model
- dense features
- pointwise scorer
↓ serve
- 6
BentoML Service
- /recommend endpoint
- A/B eval logger
Sign in to open the build guide
Free account. Get the step-by-step build and every resource link.
Take it further
- Add a real-time feature store layer using Redis so user interaction counts update without full retraining
- Implement hard-negative mining to improve tower training quality and measure the lift in recall@10
- Deploy the BentoML service to a cloud instance and add a lightweight feedback loop that logs impressions and clicks to retrigger nightly retraining
