New opportunity

ML Engineer

Catalyst Labs · New York, United States

About this role

Catalyst Labs is recruiting an ML Engineer for an unnamed New York startup whose technology analyzes real-world sales and service conversations. The role focuses on production machine learning for audio data, natural speech interaction, and real-time audio understanding. The position is based in New York, NY and is full-time. Compensation is described as an above-market base salary plus bonus and equity.

Responsibilities

Design, build, deploy, and maintain production-grade ML systems, owning the model lifecycle from data acquisition and preprocessing through training, inference, deployment, and monitoring.
Develop and optimize audio-focused models that turn unstructured voice data into business insights, and build agents that operate on real-world audio inputs.
Create scalable data pipelines for training and evaluation; improve ML infrastructure, tooling, and processes for performance and scale.
Work with cross-functional teams on the AI stack and LLM and audio ML applications. Engage directly with customers to understand needs, gather feedback, and deliver solutions.

Required qualifications

Bachelor’s or Master’s degree in Computer Science, Machine Learning, Artificial Intelligence, or a related field, and 1–6 years of professional ML engineering experience.
Strong Python skills; hands-on experience with ML frameworks such as PyTorch or TensorFlow; familiarity with cloud infrastructure; and understanding of data pipeline design, real-time inference, and model monitoring.
Experience deploying ML models to production, owning the full model lifecycle, working on audio ML or similar unstructured-data projects, and building scalable training and evaluation pipelines. Knowledge of ML systems architecture, feature engineering, evaluation strategies, and deployment best practices is expected, along with communication skills for customer and stakeholder engagement.

Preferred qualifications

TypeScript experience, AWS familiarity, and familiarity with FastAPI, OpenAI APIs, Baseten, LiteLLM, LiveKit, PostgreSQL, Redis, and S3 are pluses.

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

Skills for this role

PythonPyTorchTensorFlowAudio machine learningSpeech interactionReal-time audio understandingMachine learning model deploymentData pipelinesReal-time inferenceModel monitoringFeature engineeringModel evaluationCloud infrastructureAWSTypeScriptFastAPIOpenAI APIsBasetenLiteLLMLiveKitPostgreSQLRedisAmazon S3Customer 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 production ML end to end
The role covers data acquisition and preprocessing through training, deployment, inference, maintenance, and monitoring. Prepare a concrete example showing decisions and results across that lifecycle.
Build audio understanding and speech interaction
The job calls for audio models that extract business insights from real-world conversations, natural speech interaction, real-time understanding, and agents operating on audio inputs. Prepare to explain how you would develop and evaluate these capabilities.
Scale training, evaluation, and infrastructure
Responsibilities include scalable data pipelines and continuous improvement of ML tooling and infrastructure for performance and scale. Prepare to discuss pipeline design, evaluation strategy, and production trade-offs.
Turn customer needs into delivered solutions
The engineer works directly with customers for needs and feedback and collaborates across the AI stack. Prepare an example of translating stakeholder feedback into a technical decision and explaining its impact.

Likely challenges to think through

Balancing real-time response with model quality
Likely challenge, not a confirmed problem: speech interaction and real-time audio understanding may require latency and resource trade-offs against accuracy. Prepare to explain what you would measure and how you would choose among approaches.
Making unstructured conversations reliably useful
Likely challenge, not a confirmed problem: varied real-world audio may complicate extracting dependable, searchable business insights. Prepare to describe data quality checks, evaluation cases, and how you would investigate failures.
Maintaining models while improving the stack
Likely challenge, not a confirmed problem: end-to-end production ownership alongside infrastructure improvement can force trade-offs between immediate delivery and longer-term reliability. Prepare to explain how monitoring and evaluation would guide priorities.

Practice your answers

Technical practice

How would you take an audio ML system from concept to production?

What to coverSuggested preparation: Walk through data acquisition, preprocessing, model development, deployment, inference, and monitoring—the lifecycle named in the posting. Explain how you would make the system maintainable and decide when it is ready to release.

How would you prepare real-world conversation audio for model training?

What to coverSuggested preparation: Discuss data quality, segmentation, labeling, and handling noisy or incomplete recordings. Explain how preprocessing choices could affect the business insights extracted from voice data.

How would you choose an approach for extracting insights from sales and service conversations?

What to coverSuggested preparation: Compare plausible audio-model approaches against the insight required, available labeled data, and production constraints. Distinguish what the posting states—turning conversations into actionable data—from any specific modeling approach you propose.

What role would feature engineering play in an audio-focused ML system?

What to coverSuggested preparation: Describe how you would test whether audio-derived features add value to a model. Address trade-offs between useful signal, preprocessing cost, and reliability on real-world recordings.

How would you build scalable pipelines for training and evaluation?

What to coverSuggested preparation: Cover reproducible preprocessing, dataset versioning, and consistent training and evaluation runs. Explain how you would identify pipeline bottlenecks as data volume grows.

How would you evaluate whether an audio model produces useful business insights?

What to coverSuggested preparation: Separate model-level measures from the usefulness of outputs to sales or service teams. Discuss representative test data, error analysis, and how customer feedback could reveal failures missed by an aggregate metric.

How would you design inference for real-time audio understanding?

What to coverSuggested preparation: Explain the trade-offs among latency, accuracy, throughput, and operating cost. Describe how you would measure end-to-end responsiveness and investigate a slow stage.

How would you design an agent that operates on live audio inputs?

What to coverSuggested preparation: Outline how an agent might process incoming audio, decide when it has enough information to act, and handle uncertainty or interruption. Treat the architecture as a proposed design, not an existing company implementation.

What would you consider when deploying an audio ML service in the cloud?

What to coverSuggested preparation: Discuss serving architecture, scaling, release and rollback practices, and how you would verify production behavior. AWS and several serving tools are listed as preferences or pluses, not mandated components.

How would you monitor and improve a deployed audio model?

What to coverSuggested preparation: Cover service health, inference performance, input-data changes, and model-quality signals where outcomes are available. Explain how monitoring findings would lead to investigation, retraining, or a deployment change.

Behavioural practice

Tell me about a time you owned an ML model beyond training.

What to coverSuggested practice: Use a real situation to explain your responsibilities from data or development through deployment and maintenance. Describe your actions, the result, and what you learned about end-to-end ownership.

Tell me about a time customer feedback changed a technical solution.

What to coverSuggested practice: Describe the customer need, how you gathered or clarified feedback, and the decision you made in response. State the result without assuming the posting specifies a particular customer-engagement process.

Tell me about a time you worked across teams on an ML application.

What to coverSuggested practice: Give the situation, your role in aligning technical and nontechnical priorities, and the actions you took to resolve a disagreement or dependency. Explain the resulting decision or delivery outcome.

Tell me about a time you improved the scalability or performance of an ML workflow.

What to coverSuggested practice: Identify the bottleneck and the evidence that established it, then describe your change and its result. Focus on your contribution rather than claiming an improvement you cannot substantiate.

Tell me about a time a deployed model or ML service needed attention.

What to coverSuggested practice: If you have a relevant example, describe the production signal, how you investigated it, and the action you took. Explain the result and any monitoring or process change that followed.

Sharpen your knowledge

Topics to prepare

End-to-end ML lifecycle
Review how data acquisition, preprocessing, training, deployment, inference, and maintenance fit together. Exercise: draw a model lifecycle and identify the inputs, outputs, and failure points at each stage. The posting calls for end-to-end ownership of production ML systems.
Audio ML and speech interaction
Review how unstructured voice data can support useful predictions or insights, and what makes real-world audio difficult to process. Exercise: outline a hypothetical audio-to-insight system and explain how you would evaluate its output. This maps to the stated audio-model, natural-speech, and audio-agent responsibilities; the system design is a preparation exercise, not a specified project.
Real-time inference architecture
Review latency, throughput, reliability, and quality trade-offs for real-time audio understanding. Exercise: sketch an inference path and explain where you would measure and reduce delay. The posting explicitly expects understanding of real-time inference.
Training data and evaluation design
Review scalable pipeline design, feature engineering, and evaluation strategies for unstructured data. Exercise: define data checks, dataset splits, and metrics for a hypothetical audio model, then explain how you would avoid misleading results. These are stated areas of expected knowledge.
Production monitoring and scalability
Review how to detect changes in input data, model quality, and system performance after deployment. Exercise: propose a small set of monitoring signals and an investigation plan for a degraded audio model. The role includes production monitoring and improving ML infrastructure for scale.

Skills to revise

Python model development
Practice building and evaluating a small model on sample audio or comparable unstructured data in Python, using PyTorch or TensorFlow. Be ready to explain preprocessing, model choices, and results. Strong Python and hands-on framework experience are explicit qualifications.
Reproducible data pipeline
Build a small pipeline that ingests sample data, validates it, preprocesses it, and produces repeatable training and evaluation inputs. Explain how you would extend it for larger volumes. The posting calls for scalable training and evaluation pipelines.
Model deployment walkthrough
Practice packaging a trained model behind an inference interface and walking through versioning, deployment, and rollback decisions. The exercise is a preparation suggestion; production model deployment and lifecycle ownership are stated requirements.
Inference measurement and monitoring
Run a small inference test, measure response time and failures, and define checks for input and output quality. Explain what you would investigate if performance worsened. This supports the stated real-time inference and model-monitoring responsibilities.
Customer-needs conversation
Rehearse asking a customer what insight they need from conversations, clarifying how they would use it, and summarizing the request as a testable ML objective. This is practice for the role's direct customer engagement, not a claim about a specific interview exercise.

Your preparation checklist

Prepare a production ML walkthrough
Choose a real project and make a one-page walkthrough of data acquisition, preprocessing, training, deployment, inference, monitoring, and maintenance. The role calls for end-to-end ownership; distinguish what you personally built from team contributions.
Show relevant audio or unstructured-data work
Prepare an example involving audio ML or a similar unstructured-data problem, as the posting allows. Explain the input, model choices, feature engineering, evaluation, and how the output was used.
Practice an audio-system design exercise
Sketch a proposed system for extracting insights from real-world conversations. Cover data flow, scalable training and evaluation pipelines, real-time inference, and monitoring; label design choices as proposals, not details of the employer’s system.
Explain a speech-interaction approach
Outline how you would approach natural speech interaction or an agent operating on live audio. Be ready to discuss latency, evaluation, and failure handling without assuming a particular product architecture.
Gather evidence for the technical requirements
Prepare concise examples of Python, PyTorch or TensorFlow, cloud infrastructure, and production deployment work. Review any relevant experience with the listed pluses, such as AWS or FastAPI, without presenting them as required.
Rehearse a customer-facing discussion
Prepare an example of gathering stakeholder needs, incorporating feedback, and explaining an ML trade-off. The role explicitly includes direct customer engagement and cross-functional work.
Bring questions about the actual role
Ask which audio tasks matter most, how success is evaluated, what real-time constraints apply, and how model ownership and customer feedback work. These clarify responsibilities the posting describes without assuming how the unnamed startup operates.

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…