Recommendation Engine Architecture
The full recommendation pipeline for a streaming platform covering four kinds of content — film, music, podcasts, live events. Narrow millions of items down to a few hundred candidates, rank those, then arrange them into rows under business rules. Includes the policy that decides how much a skip counts against an item compared to a full play.
Spec · ~50pp
Synthetic User Population Simulator
A simulated audience for testing recommendations before launch. Built on OASIS, with taste profiles written by a 235B-parameter model, four types of user behavior, a model of how people drift away over time, and checks that the fake population behaves like a real one. Lets you measure a recommendation system before you have any users.
Design · Agent Sim
Cross-Domain Embedding Space
One shared numeric representation for every kind of content, so a podcast episode and a concert ticket can be compared directly. Same encoder and same description template for all four content types, with per-type bias removed afterward and the result verified by a probe that checks the space still carries real meaning.
Design · Embeddings
Two-Tower Retrieval Model
Two neural networks trained together — one that encodes a user, one that encodes an item — so finding matches is a fast nearest-neighbor lookup at serving time. Trained with item IDs randomly hidden, so the model can still rank something it has never seen before, which is the normal case for a catalog that adds new titles every week.
Design · Retrieval
Event Taxonomy & Data Architecture
How every play, skip, and click gets recorded and stored — ClickHouse tables fed by a Kafka/Redpanda stream, with the schema designed so the data you’ll want in eighteen months is already being captured now.
Design · ClickHouse
Privacy & Compliance Framework
The legal side of running a recommender in the EU: the formal privacy risk assessment GDPR requires, deletion handled by destroying encryption keys so the data becomes unreadable, event capture that only fires if the user consented, and the explanation of why a user is seeing what they’re seeing that EU law now requires.
Design · GDPR / DSA
ML Org Infrastructure
Where the models and training data live and how they’re versioned — a HuggingFace registry plus small internal web apps so QA can poke at a model in a browser instead of running a notebook.
Design · HF / Gradio