New opportunity

Member of Technical Staff - Applied ML, RecSys

Liquid AI · Boston, United States

About this role

Liquid AI is hiring a Member of Technical Staff - Applied ML, RecSys to own applied machine learning for enterprise recommendation and ranking workloads. The role adapts Liquid Foundation Models for customer-specific personalization under production constraints, with responsibility from requirements through delivery and evaluation. The posting lists Boston as the location and Hybrid as the work arrangement.

Responsibilities

Serve as technical owner for enterprise customer engagements, translating business requirements and metrics into recommendation model specifications and technical decisions.
Design and execute large-scale pipelines for user interaction data, feature engineering and training data curation. Fine-tune and adapt sequential recommendation models, including HSTU-style architectures, for customer use cases.
Design and interpret task-specific evaluations covering ranking quality, latency and throughput. Make tradeoffs among model quality, production performance and measurable outcomes such as engagement, conversion and revenue lift.
Build reusable applied tooling and workflows between engagements. In the first year, independently deliver customer engagements with minimal oversight and become a trusted technical owner.

Required experience:

Hands-on experience building or fine-tuning recommendation models at scale beyond off-the-shelf collaborative filtering; experience with sequential recommendation architectures, user behavior modeling or large-scale ranking systems.
Strong judgment about recommendation data quality and evaluation, including offline metrics, A/B testing and alignment with business metrics; experience with large-scale interaction-data pipelines and feature engineering.
Proficiency in Python and PyTorch, autonomous coding and debugging, ownership of customer work, and clear communication between business and technical stakeholders.

Preferred experience includes transformer-based recommendation architectures such as HSTU, SASRec or BERT4Rec; delivering recommendation systems to external customers with measurable results; and serving models under latency and throughput constraints.

Benefits include a competitive base salary with equity, fully paid medical, dental and vision premiums for employees and dependents, 401(k) matching up to 4% of base pay, unlimited PTO and company-wide Refill Days. No salary amount or minimum years of experience is specified.

For full details about this role, click “Apply on company site” to view the employer’s original posting.

Skills for this role

Recommendation systemsRanking systemsSequential recommendationUser behavior modelingFeature engineeringData pipelinesTraining data curationModel fine-tuningModel evaluationOffline metricsA/B testingPythonPyTorchHSTUSASRecBERT4RecTechnical communication

Checking your profile…

Preparation suggestions, not the employer’s actual interview questions. Always check the original posting for current requirements.

Understand the role

What matters most

Own customer engagements end to end
The role is the technical owner from requirements through delivery and evaluation, translating customer business metrics into recommendation specifications and technical decisions. Year-one success includes delivering engagements independently with minimal oversight.
Build reliable interaction-data foundations
The work includes large-scale user interaction pipelines, feature engineering, and training-data curation. The posting emphasizes judgment about data quality for recommendation systems.
Adapt sequential recommendation models
The role fine-tunes and adapts large-scale sequential models, including HSTU-style architectures, for customer-specific personalization and ranking. Hands-on recommendation modeling at scale is a must-have; specific transformer architectures are nice-to-have.
Evaluate for quality and business impact
Task-specific evaluation must cover ranking quality, latency, and throughput. The posting also calls for offline metrics, A/B testing, and alignment with outcomes such as engagement, conversion, and revenue lift.
Create reusable delivery workflows
Between customer engagements, the role builds applied tooling and workflows that accelerate future work; this is also an explicit year-one success measure.

Likely challenges to think through

Turning business goals into model decisions
Preparation hypothesis: customer metrics may not map cleanly to a ranking objective, creating tension between a technically attractive model and a measurable outcome. Prepare to explain how you would define a model specification, choose aligned evaluation metrics, and push back on a poorly matched objective.
Balancing ranking gains against production cost
Preparation hypothesis: adapting a larger sequential model may improve ranking quality while increasing latency or reducing throughput. Prepare to describe how you would compare model variants using task-specific quality and performance results, then justify a customer-facing trade-off.
Reusing work without flattening customer differences
Preparation hypothesis: customer-specific data and goals may require different adaptations, while the role also calls for reusable workflows. Prepare to distinguish components you would standardize, such as evaluation or data-validation steps, from decisions that must remain specific to an engagement.

Practice your answers

Technical practice

How would you turn an enterprise customer's personalization goal into a recommendation model specification?

What to coverThe role includes translating customer requirements into model specifications. Suggested preparation: Explain how you would define the ranking task, available interaction data, success metrics, and production constraints before choosing a modeling approach.

How would you design a large-scale user-interaction pipeline for recommendation training?

What to coverThe posting calls for interaction-data pipelines and feature engineering at scale. Suggested preparation: Discuss event definitions, sequence construction, feature consistency, and checks that make the resulting training data usable.

What data-quality problems would you investigate before fine-tuning a recommender?

What to coverData quality and training-data curation are explicit parts of the work. Suggested preparation: Describe how you would detect missing or duplicated events, unreliable labels, and leakage, then assess their effect on recommendations.

When would sequential user-behavior modeling be preferable to simpler recommendation approaches?

What to coverThe role calls for experience beyond off-the-shelf collaborative filtering and includes sequential models. Suggested preparation: Explain what ordering and recency can contribute, and weigh that value against data, training, and serving costs.

How would you adapt a large sequential recommendation model to a customer-specific ranking task?

What to coverFine-tuning and adapting models, including HSTU-style architectures, is stated work; HSTU, SASRec, and BERT4Rec experience is listed as preferred. Suggested preparation: Walk through your choice of training objective, fine-tuning scope, and comparisons that would show whether adaptation helps.

How would you design an offline evaluation for a customer-specific ranking model?

What to coverThe posting asks for task-specific evaluation and sound judgment about offline metrics. Suggested preparation: Explain how you would choose ranking metrics and evaluation data, avoid temporal leakage, and interpret results for the customer's task.

How would you test whether better offline ranking translates into business impact?

What to coverA/B testing and alignment with engagement, conversion, or revenue lift are named in the posting. Suggested preparation: Discuss a suitable experiment comparison, the primary business metric, and how you would interpret conflicting offline and online results.

How would you choose between ranking quality and serving latency or throughput?

What to coverProduction performance and trade-offs among model quality, latency, throughput, and business impact are central to the role. Suggested preparation: Describe how you would measure these dimensions together and decide whether a quality gain justifies its serving cost.

How would you debug a PyTorch recommendation model whose training results are unexpectedly poor?

What to coverThe posting requires Python and PyTorch proficiency with autonomous coding and debugging. Suggested preparation: Lay out a systematic check of inputs, sequence and label construction, training behavior, and evaluation before changing the architecture.

What would you make reusable across distinct customer recommendation engagements?

What to coverBuilding applied tooling and workflows that accelerate future engagements is an explicit responsibility. Suggested preparation: Identify repeatable parts of data validation, training, or evaluation while explaining how customer-specific requirements would remain configurable.

Behavioural practice

Tell me about a time you owned an applied ML engagement from requirements through evaluation.

What to coverEnd-to-end customer ownership with minimal oversight is a stated expectation. Suggested preparation: Use a real situation, your decisions and actions, and the delivered result; distinguish what you owned from others' contributions.

Describe a time you translated a business objective into technical decisions.

What to coverThe role requires communication between customer business metrics and technical choices. Suggested preparation: Give a real situation, explain how you clarified the objective and communicated your actions, and state the resulting decision or outcome.

Tell me about a time you pushed back on a stakeholder's requested approach.

What to coverThe posting explicitly values clear communication and pushing back when needed. Suggested preparation: Describe the situation, the evidence and alternative you presented, and the result without treating disagreement as the goal.

Describe a time you chose a pragmatic solution over a more complex model.

What to coverThe role prioritizes measurable outcomes and judgment about quality, latency, and business impact. Suggested preparation: Use a real situation to explain the competing options, your decision and actions, and what the results showed.

Tell me about a workflow or tool you made reusable for later work.

What to coverReusable applied tooling across engagements is part of the stated first-year success criteria. Suggested preparation: Describe the recurring problem, what you built or changed, and evidence of its effect on subsequent delivery.

Sharpen your knowledge

Topics to prepare

Sequential recommendation and user behavior
The posting requires experience with sequential architectures, user behavior modeling, or large-scale ranking beyond off-the-shelf collaborative filtering. Suggested exercise: explain how interaction order could change a ranking decision, then compare a sequential approach with collaborative filtering. HSTU, SASRec, and BERT4Rec are listed as nice-to-have examples, not required architectures.
Interaction-data quality and training-data design
The role calls for judgment about data quality, feature engineering, and training-data curation at scale. Suggested exercise: sketch a chronological training dataset from user interactions and identify missing events, duplicate events, and features that would not be available at prediction time.
Recommendation evaluation and business alignment
The posting emphasizes offline metrics, A/B testing, and measurable outcomes such as engagement, conversion, and revenue lift. Suggested exercise: choose an offline ranking metric for a hypothetical customer goal, explain its limitations, and outline what an A/B test would need to measure.
Quality versus production performance
Task-specific evaluation must cover ranking quality, latency, and throughput; year-one success includes judging tradeoffs against business impact. Suggested exercise: compare two hypothetical models where one ranks better but responds more slowly, and state what evidence would guide a delivery decision.
Customer requirements to model specifications
The technical owner translates enterprise customer requirements into recommendation specifications and communicates decisions to business and technical stakeholders. Suggested exercise: turn a hypothetical personalization goal into a proposed ranking task, success measure, and production constraint, clearly labeling any assumptions.

Skills to revise

Build and debug a model adaptation in Python and PyTorch
Python, PyTorch, autonomous debugging, and building or fine-tuning recommendation models are explicit must-haves. Suggested exercise: implement a small sequence-based ranking training loop, debug a deliberately incorrect input shape or label alignment, and explain what you checked. This is practice, not a prescribed employer project.
Construct an interaction-data pipeline
The work includes executing large-scale pipelines for interaction data, feature engineering, and training-data curation. Suggested exercise: write a Python workflow that orders sample events by user and time, produces training examples and features, and checks counts and chronology before training.
Run and interpret task-specific evaluations
The role requires evaluating ranking quality, latency, and throughput. Suggested exercise: evaluate two sample rankers on held-out interactions, measure their response times and throughput, and write a short decision note that distinguishes observed results from anticipated business outcomes.
Present a customer-facing technical decision
End-to-end customer ownership, clear stakeholder communication, and reusable applied workflows are part of the posting. Suggested exercise: draft a brief model specification and evaluation summary for a hypothetical engagement, explain a quality-versus-latency tradeoff in business terms, and turn the repeatable steps into a checklist.

Your preparation checklist

Map a customer goal to a model specification
The role owns enterprise engagements from requirements through evaluation. Prepare a one-page example showing how an engagement or hypothetical brief becomes a ranking objective, measurable business metric, technical specification and delivery plan.
Walk through an interaction-data pipeline
Prepare a diagram of a pipeline you built, or a clearly labeled design exercise, covering user events, data-quality checks, feature engineering and training-data curation at scale. Be ready to explain consequential data decisions.
Explain a sequential recommendation approach
The posting calls for recommendation modeling beyond off-the-shelf collaborative filtering. Prepare a project walkthrough or design comparison showing why you would use a sequential model, how you would adapt it to a customer use case and where Python and PyTorch fit. HSTU, SASRec and BERT4Rec are listed as nice-to-have examples.
Draft an evaluation plan
Produce a compact plan connecting offline ranking metrics, an A/B test and business outcomes such as engagement, conversion or revenue lift. Include how you would interpret a quality gain that does not translate into customer impact.
Practice a production tradeoff exercise
Sketch how you would compare candidate models on ranking quality, latency and throughput, then explain a defensible deployment choice under hypothetical constraints. The posting emphasizes judgment about these tradeoffs, not a particular serving stack.
Prepare evidence of independent delivery
Select a real example, if available, that shows autonomous Python/PyTorch debugging, technical ownership, stakeholder communication and a reusable workflow or tool. Bring a concise account of your decisions, results and what you would change; do not claim outcomes you cannot substantiate.
Write questions for the employer
Prepare questions about how customer business metrics become model specifications, what evaluation and production constraints typically shape engagements, and where reusable applied tooling would have the most value.

Availability last checked: Nov 4, 2026. Status can change after a check; confirm on the employer’s website. Report an issue · Our checking policy

KEEP EXPLORING

Similar AI jobs

Related skills and specializations in United States. Matched to this role, not your profile.

Explore more jobs

Finding similar opportunities…