Experiment tracking, data versioning, model registry
Vector Similarity Search
What sets us apart
We base our recommendations solely on user watch history, allowing us to train on unrelated datasets.
ENHANCED WITH LLM Embeddings
We use OpenAI’s API to learn embeddings from movie plot and other textual features, reducing overfitting, enhancing model results and enabling the model to recommend new or lesser-known movies.
IMPROVED Embedding Techniques
By concatenating ratings into user watch history before applying an attention layer, we supercharge our model metrics.
Best Practices in ML Operations
Key Steps for Effective Model Deployment and Performance Optimization
We set budget-aligned limits on deployment from the start, avoiding unnecessary investments in design and models, quickly abandoning frameworks and models that would waste time.
Using ClearML, we document training runs with code changes, dataset versions, and hyperparams for easy result comparison and reproduction.
Our system is continuously monitored with Prometheus and Grafana, ensuring production success. We define metrics to compare the live model to the offline evaluations, and measure speed, throughput and other metrics.