Description

IMDb is the world's most popular and trusted source for information on movies, TV shows, and celebrities. Products and services to help fans discover and decide what to watch and where to watch it include: the IMDb website for desktop and mobile devices; apps for iOS and Android; and X-Ray on Prime Video. For entertainment industry professionals, IMDb provides IMDbPro and Box Office Mojo. IMDb licenses information from its vast and trusted database to third-party businesses worldwide. As an Amazon company, IMDb employees enjoy the benefits and resources of a tech giant, while maintaining the autonomy and impact of a small, nimble team.

Every product decision at IMDb surfaces a question the data must answer. The IMDb data science team's work spans human versus automated traffic detection, moderation classification, abuse modeling, audience demand forecasting, advertising enhancements, content understanding; and the scope continues to expand.

We are hiring a data scientist who wants that breadth. You will frame loosely defined business problems as well-specified science problems and select the method that fits the data and the question rather than the one you know best. Measurement ownership runs end to end: ground truth design, evaluation, and proof that the model moved the metric it was built to move.

We expect you to use frontier models to scale your own throughput, not only to embed them in products; fluency with these tools is part of the craft. Success in this position demands exceptional scientific leadership, strong organizational skills, software engineering excellence, product vision, and sharp business acumen, along with expertise in implementing responsible AI solutions.

Key job responsibilities
• Frame loosely defined business problems as science problems, and iterate on the approach with limited guidance.
• Establish ground truth. Design the sampling, set the label quality bar, and build datasets across behavioral, text, and catalog domains.
• Define precision, recall, and coverage, and report what accuracy is achievable at what cost.
• Document failure modes, quantify the cost of each error type, and set thresholds with product and policy partners before deployment.
• Deliver methodology and analysis on medium-size projects, written so other scientists and business partners can both act on the results.
• Decide whether to train in-house, fine-tune, or call a language model, and make that call on evidence.
• Use generative AI tooling to scale your own throughput rather than only to build products.
• Ship solutions another scientist can test and reproduce.
• Partner with engineers to move models into production, then define drift monitoring and retraining triggers.
• Build deep knowledge of what the team owns, and mentor teammates on method choice.

A day in the life
Most of the week is modeling and measurement: pulling data, building features, running evaluations, and discussing thresholds with evidence rather than opinion. You use AI tooling to handle the repetitive parts, which keeps your focus on the actual science. The rest of your time goes to the product managers, policy partners, and operations team who act on your model's outputs every day. Because the team is small, you move between problem areas rather than owning just one.