New opportunity

ML Engineer

Catalyst Labs · New York, United States

About this role

Catalyst Labs is recruiting an ML Engineer for a VC-backed startup whose AI captures and analyzes field sales and service conversations. The role focuses on production ML systems for natural speech interaction, real-time audio understanding, and insights from unstructured voice data. The posting header lists Boston, Massachusetts, United States, while the role details state “Location: New York, NY”; no working arrangement is specified. Compensation is described as above-market base pay, bonus, and equity.

Responsibilities

Design, build, deploy, and maintain production-grade ML systems, owning the lifecycle from data acquisition and preprocessing through training, inference, deployment, and monitoring.
Develop and optimize audio-focused models, build agents that operate on real-world audio, and deliver AI-powered speech solutions.
Build scalable pipelines for model training and evaluation; improve ML infrastructure, tooling, scalability, and performance.
Collaborate across teams on the AI stack and LLM and audio ML applications. Work directly with customers to identify 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 and hands-on experience with ML frameworks such as PyTorch or TensorFlow. Experience deploying models to production and owning the full model lifecycle.
Experience with audio-focused ML or comparable unstructured-data projects; scalable data pipelines; and ML systems architecture, feature engineering, evaluation, and deployment practices.
Familiarity with cloud environments, real-time inference, and model monitoring, plus strong communication skills for customer and stakeholder engagement.

Preferred

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

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

Skills for this role

PythonTypeScriptPyTorchTensorFlowAWSData pipelinesReal-time inferenceModel monitoringAudio machine learningAI agentsLarge language modelsFastAPIOpenAI APIsBasetenLiteLLMLiveKitPostgreSQLRedisAmazon S3Feature engineeringModel evaluationCloud infrastructureCustomer 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, preprocessing, training, deployment, inference, and monitoring. Prepare to explain how you would make a model reliable after launch, not just during development.
Build speech and audio capabilities
The posting emphasizes natural speech interaction, real-time audio understanding, audio-focused models, and agents operating on real-world audio. Focus preparation on technical decisions specific to voice inputs and outputs.
Turn voice data into useful insights
The client’s product analyzes field sales and service conversations, and the role calls for extracting business-critical insights from unstructured voice data. Be ready to discuss how model outputs could be evaluated for usefulness, without assuming a particular product feature.
Scale training and evaluation infrastructure
Responsibilities include scalable pipelines, better tooling, and improved ML infrastructure performance. Prepare an example of how pipeline or architecture choices affect repeatable training and evaluation.
Work across teams and with customers
The engineer will help shape the AI stack with cross-functional teams and work directly with customers to identify needs and gather feedback. Prepare to explain how feedback can inform technical priorities.

Likely challenges to think through

Balance response time and audio quality
Likely preparation hypothesis: real-time audio understanding may require trading model complexity against inference latency. Prepare to describe how you would measure both and decide when an optimization is acceptable.
Evaluate variable real-world conversations
Likely preparation hypothesis: unstructured field audio may be harder to evaluate consistently than controlled inputs, while the output still needs to yield useful insights. Prepare a validation approach that accounts for variation in audio and checks whether extracted information serves the intended use.
Balance customer needs with a scalable system
Likely preparation hypothesis: direct customer feedback may point to specific needs, while the role also calls for scalable pipelines and infrastructure. Prepare to explain how you would prioritize a requested improvement and determine whether it belongs in the shared ML system.

Practice your answers

Technical practice

How would you acquire and preprocess real-world conversation audio for an ML system?

What to coverThe role covers data acquisition and preprocessing for audio-focused ML. Suggested preparation: discuss audio quality, segmentation, labeling, and how you would keep training data representative of field conversations.

How would you develop and optimize a model for understanding speech in field conversations?

What to coverThe posting calls for audio-focused models and hands-on experience with frameworks such as PyTorch or TensorFlow. Suggested preparation: explain model selection, training, and the trade-offs between accuracy, compute cost, and inference speed.

How would you evaluate whether a model extracts useful insights from voice data?

What to coverThe role aims to turn unstructured conversations into actionable information and calls for evaluation strategies. Suggested preparation: define task-specific measures, examine errors against the intended customer use, and explain how you would compare model versions.

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

What to coverReal-time audio understanding and inference are stated parts of the role. Suggested preparation: describe streaming input, latency and throughput constraints, and how you would handle partial or delayed results.

How would you design an agent that operates on real-world audio inputs?

What to coverBuilding agents that operate natively on audio is an explicit responsibility. Suggested preparation: explain how audio understanding would inform the agent’s next action, and how you would test behavior when the input is ambiguous.

How would you approach an AI-powered natural speech interaction?

What to coverThe posting calls for speech solutions and collaboration on LLM and audio ML applications. Suggested preparation: describe how speech input, language-model behavior, and spoken responses could fit together, including latency and error-propagation trade-offs.

How would you build a scalable pipeline for training and evaluating audio models?

What to coverScalable data pipelines for model training and evaluation are part of the stated work. Suggested preparation: cover repeatable preprocessing, dataset versioning, reproducible evaluations, and where pipeline bottlenecks might arise.

How would you take an audio model from development into production?

What to coverThe role requires ownership from model development through deployment and maintenance, with cloud familiarity preferred. Suggested preparation: outline a deployment approach, validation before release, and how you would manage model and service changes without assuming a particular platform.

How would you diagnose a production model whose results have deteriorated?

What to coverProduction monitoring is explicitly included in the model lifecycle. Suggested preparation: explain which input, output, and service signals you would inspect, and how you would distinguish data changes from model or infrastructure issues.

How would you improve the scalability and performance of an existing ML system?

What to coverThe posting calls for continuous improvement of ML infrastructure and processes. Suggested preparation: describe how you would locate a bottleneck, compare possible changes, and check that performance gains do not undermine model quality or reliability.

Behavioural practice

When have you owned an ML project across multiple lifecycle stages?

What to coverEnd-to-end ownership is central to the role. Suggested practice: give a real situation, the stages you personally handled and decisions you made, and the resulting outcome.

When has customer feedback changed a technical solution you were building?

What to coverThe role involves working directly with customers to identify needs and gather feedback. Suggested practice: describe the customer need, how you acted on the feedback, and what changed as a result.

When have you worked across teams to resolve a disagreement about an ML approach?

What to coverCross-functional collaboration on the AI stack is stated in the posting. Suggested practice: explain the situation, how you communicated the trade-offs and reached a decision, and the result.

When have you improved an ML workflow or system after identifying a recurring problem?

What to coverThe role includes improving tooling, infrastructure, scalability, and performance. Suggested practice: identify the problem and your specific action, then describe the observed result without overstating your contribution.

Sharpen your knowledge

Topics to prepare

Audio understanding and speech interaction
Revision exercise: Sketch how a system could turn noisy field conversations into searchable insights, and identify where speech understanding might fail. Relevance: the role explicitly focuses on natural speech interaction, real-time audio understanding, and insights from unstructured voice data.
Real-time ML architecture
Revision exercise: Diagram an audio inference path and explain trade-offs among latency, throughput, and reliability. Relevance: real-time inference, scalability, and ML systems architecture are named in the posting.
Model evaluation and feature engineering
Revision exercise: Define evaluation measures for an audio model and explain how you would detect whether its outputs remain useful on new conversations. Relevance: the posting calls for feature engineering, evaluation strategies, and models that extract business-critical insights.
Production model lifecycle
Revision exercise: Walk through decisions from data acquisition and preprocessing to training, deployment, inference, and monitoring. Relevance: end-to-end ownership and production maintenance are explicit responsibilities.
Agents working with audio
Revision exercise: Describe an agent’s input, decisions, outputs, and failure handling when operating on real-world audio. Relevance: building audio-native agents and contributing to LLM and audio ML applications are stated responsibilities.

Skills to revise

Python model implementation
Revision exercise: Implement and test a small audio-model training and inference loop in Python using PyTorch or TensorFlow. Relevance: strong Python skills and hands-on experience with one of these ML frameworks are required.
Scalable training and evaluation pipelines
Revision exercise: Build a reproducible pipeline that ingests example audio data, preprocesses it, trains a model, and produces evaluation results; note where scaling would change the design. Relevance: scalable pipelines for model training and evaluation are core experience in the posting.
Deployment and monitoring
Revision exercise: Package a model for inference, define a deployment check, and specify what you would monitor for performance and failures. Relevance: production deployment, maintenance, and model monitoring are explicit expectations.
Customer-focused technical communication
Revision exercise: Practice turning a hypothetical customer request for conversation insights into a concise problem statement, success measure, and feedback questions. Relevance: the engineer will work directly with customers to identify needs, gather feedback, and deliver solutions.

Your preparation checklist

Prepare a production ML walkthrough
Choose a project you can discuss from data acquisition and preprocessing through training, deployment, inference, and monitoring. Bring a diagram, your specific decisions, and measured outcomes where available; the posting emphasizes end-to-end ownership.
Explain an audio or unstructured-data project
Prepare a short account of the input data, model approach, evaluation, and limitations. Audio-focused ML or comparable unstructured-data experience is listed; do not present a comparable project as audio work.
Sketch a real-time speech system
As a practice exercise, diagram how field-conversation audio could become searchable insights or support an audio agent. Explain inference latency, failure handling, scaling, and monitoring as proposed design choices, not as claims about the employer’s architecture.
Show pipeline and implementation evidence
Identify concrete examples of Python and PyTorch or TensorFlow work, plus a training or evaluation pipeline you helped build. Be ready to explain data handling, feature engineering, reproducibility, and deployment trade-offs.
Rehearse a customer-facing example
Prepare one example of gathering stakeholder needs, incorporating feedback, and explaining a technical trade-off. Direct customer engagement and cross-functional collaboration are part of the role.
Draft targeted questions
Ask how the client evaluates speech quality and real-time performance, what parts of the ML lifecycle this hire would own, and which listed tools are actually used. Clarify the location and working arrangement: the page header says Boston, the role details say New York, and no arrangement is specified.

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…