Why this role is interesting (if you're an ML-engineering person)

‍Most ML jobs in e-commerce are interchangeable. This one isn't:

• The catalogue is the product, and it's visual. Computer vision, embeddings, and multimodal LLMs aren’t bolted to a 10-year-old PHP monolith. Visual similarity, image moderation, and image-to-metadata get invested in, not tolerated.
• Adversarial, high-stakes, and very real. IP/copyright detection isn't a hypothetical "nice to have" - it's daily work with Legal and BizDev on Disney, Marvel, Warhammer, CD Projekt, Square Enix, Nintendo, NBA, and dozens more. AI-art detection is a live policy debate. Artist quality scoring decides whether the marketplace can scale 10x.
• You ship, you own. No separate ML-ops team, no six layers of approval. You take a model from notebook to Vertex AI / k8s / AWS, monitor it, and iterate - in days, not quarters.

‍What you'll work on

• Art Taxonomy & image understanding. The internal name for image-to-metadata at Displate - using multimodal LLMs (Gemini via Vertex AI). Feeds search, discovery, moderation, and analytics.
• Search ranking & retrieval (Search v3). ELSER, building the search-arena offline evaluation pipeline (BigQuery → GCS → ES eval), running A/B tests against real revenue, and shipping rescoring scripts that actually move conversion.
• IP & copyright detection at scale. Recurring, business-critical, often Slack-driven work with Legal and BizDev - finding and removing infringing artworks across hundreds of IPs, building automated pipelines that turn a "we just signed a deal with Studio X" message into a clean catalogue. Includes the marketplace-reopening initiative (Project Behemoth).
• Image Quality Assessment - especially for artist uploaders, to make sure what they provide is not blurred, printed out wrong, low quality, not compliant with Terms of Service, etc.
• Recommendations, customer & sales ML. Visual similarity (CLIP / DINO / SigLIP / open to others)
• Dynamic User Preference Engine:  Build a "living" user preference system that tracks and models behavioral data in real-time.
• Construct an Artwork Trend Graph: Develop a comprehensive semantic graph representation of our entire artwork catalog, mapping out complex stylistic relationships and intellectual property connections, to spot viral market momentum.

What we're looking for
Must-haves (3+ years):

• You need to have a knack for connecting the dots between various data sources - you won't limit yourself to ML tables, but talk to other teams to understand data collection, its ETL, and flow in order to conceptualize how to make use of it in a combined manner
• Strong Python and the modern data-science stack (PyTorch, HuggingFace).
• Experience deploying ML models to production on GCP and/or AWS (Vertex AI, Cloud Run, Kubernetes, AWS Personalize, or equivalent).
• LLMs in production (Gemini, GPT, Claude) with Langfuse / Pydantic AI / structured output.
• Building and maintaining ML APIs (FastAPI).
• Working knowledge of embeddings, CNNs, and transformers -- with shipped experience in at least one of: visual similarity, recommendations, search ranking, or computer vision.
• Comfort with BigQuery and SQL (Postgres / MySQL).
• Solid software-engineering fundamentals: Git, code review, testing, CI/CD.

‍Nice-to-haves:

• Elasticsearch or other modern search/ranking stacks.
• Generative image work (Stable Diffusion, Flux, Nano Banana, ComfyUI, LoRA fine-tuning).
• Tabular ML and explainability (LightGBM / XGBoost, SHAP, MLflow, Optuna).
• Bayesian modelling (PyMC, BG/NBD-style customer models).
• Content moderation, copyright/trademarked detection.
• E-commerce recommendations for AI/ML features.
• Some JAVA proficiency (so you can eavesdrop on engineers roasting others during PR)