Senior ML Encoder Lead

About the role

Build the first shared learned representation of our customers — one dense vector per customer, trained on longitudinal transaction, sales and interaction history — that downstream GenAI and analytics products can reuse instead of each re-deriving its own view of the same market.

The contractor will design the pretraining objective, train and evaluate the encoder, and produce the evidence that determines whether the approach continues. Evaluation is as much of the deliverable as the model.

This is a hands-on senior contractor role requiring the ability to define modeling objectives and evaluation design, and write production code — not execute a predefined specification.

Location: South San Francisco, CA (3 days onsite/week). Duration: 6+ months.

Responsibilities

• Design the pretraining objective for a customer representation encoder.
• Train and evaluate the encoder model, ensuring rigorous assessment of its effectiveness.
• Produce evidence to determine the viability of the approach for continued development.
• Write production-grade code for model development, training, and evaluation.
• Carry models from research into production, including data contracts, training pipelines, versioning, serving, monitoring, and reproducibility.
• Present findings, uncertainty, and recommendations to senior stakeholders, including stopping non-viable approaches.

Requirements

• Proven experience in representation learning and encoder development, including designing pretraining objectives — not just consuming pre-trained embeddings or fine-tuning published models.
• Deep expertise in self-supervised or contrastive pretraining, sequence and temporal modeling, transformers, graph neural networks, or recommender embeddings.
• Experience modeling large, sparse, longitudinal event data such as transactions, claims, clickstream, customer journeys, or engagement histories.
• Experience building inductive representations to handle entities with little history, avoiding reliance on lookup tables.
• Rigorous evaluation practices, including time-based splits, leakage detection, cold-start slices, transfer to held-out populations, uncertainty quantification, and hard baselines.
• Ability to assess whether an embedding provides genuine incremental signal downstream, including calibration, stability, drift, and subgroup performance.
• Strong Python engineering skills with PyTorch or JAX, SQL, and distributed data processing for large-scale cloud-based model training.
• Experience taking models from research to production, including data contracts, training pipelines, versioning, serving, and monitoring.

Preferred Qualifications

• Experience with customer-360 representations, behavioral embeddings, recommender systems, or foundation models over event data.
• Familiarity with privacy, fairness, and re-identification risks in learned representations of individuals.
• Publications, patents, or public applied work in representation learning.
• Experience in industries with large-scale behavioral event data, such as consumer technology, marketplaces, streaming, financial services, payments, or advertising technology.