New opportunity

ML Engineer

Catalyst Labs · New York, United States

About this role

Catalyst Labs is recruiting an ML Engineer for a New York-based startup whose AI technology captures and analyzes real-world conversations for outside sales and service teams. The role focuses on production ML systems that turn audio and other unstructured voice data into searchable, actionable business insights. 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 lifecycle from data acquisition and preprocessing through training, inference, deployment, and monitoring.
Develop and optimize audio-focused models, natural speech interactions, real-time audio understanding, and agents that operate on real-world audio inputs.
Build scalable data pipelines for model training and evaluation; improve ML infrastructure, tooling, performance, and scalability.
Collaborate across teams on the AI stack and LLM and audio ML applications. Work 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 programming skills; hands-on experience with frameworks such as PyTorch or TensorFlow; familiarity with cloud infrastructure, data pipeline design, real-time inference, and model monitoring.
Experience deploying ML models to production and owning the full model lifecycle. Experience with audio-focused ML or similar unstructured-data projects, scalable training and evaluation pipelines, ML systems architecture, feature engineering, evaluation strategies, and deployment practices.
Communication skills for direct engagement with customers and stakeholders.

Preferred experience: TypeScript and AWS; familiarity with FastAPI, OpenAI APIs, Baseten, LiteLLM, LiveKit, PostgreSQL, Redis, and S3 is a plus.

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

Skills for this role

PythonTypeScriptPyTorchTensorFlowAWSData pipeline designReal-time inferenceModel monitoringAudio machine learningAI agentsLarge language modelsFastAPIOpenAI APIsBasetenLiteLLMLiveKitPostgreSQLRedisAmazon S3Feature engineeringModel evaluationML systems architectureCustomer 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 explicitly covers data acquisition, preprocessing, training, deployment, inference, maintenance, and monitoring. Prepare to explain how you would make a model reliable after launch, not just train it.
Turn audio into usable insights
The client wants models for unstructured voice data, natural speech interaction, real-time audio understanding, and agents operating on audio inputs. Focus preparation on how audio outputs become searchable, actionable information for sales and service teams.
Scale training, evaluation, and infrastructure
The posting calls for scalable data pipelines, evaluation strategies, and improvements to ML tooling and performance. Prepare to discuss design choices and how you would measure whether a change improves the system.
Work across teams and with customers
The engineer will help shape the AI stack with cross-functional teams and speak directly with customers about needs and feedback. Prepare examples of translating feedback into a technical decision and communicating its trade-offs.

Likely challenges to think through

Balance audio quality with response time
A likely challenge, not a confirmed problem, is balancing accurate understanding of real-world audio against the latency of real-time interaction. Prepare to discuss how you would test both quality and inference speed before choosing a design.
Evaluate insights, not only model outputs
Because the stated goal is actionable insight from conversations, strong model metrics may not by themselves show business usefulness. Prepare a proposed evaluation approach that connects model results to customer feedback, while keeping those measures distinct.
Balance customer requests with scalable systems
Direct customer work alongside infrastructure ownership may create tension between a specific request and a reusable solution. Prepare to explain how you would clarify the need, scope an initial delivery, and decide what belongs in shared ML infrastructure.

Practice your answers

Technical practice

How would you take an audio ML system from data acquisition through production maintenance?

What to coverThe posting calls for ownership across acquisition, preprocessing, training, deployment, inference, and monitoring. Suggested preparation: explain the decisions and checks at each stage, including how production observations would inform later model changes.

How would you prepare real-world voice data for model training?

What to coverThe role involves audio-focused ML and unstructured voice data. Suggested preparation: discuss data-quality checks, preprocessing and feature-engineering choices, and how you would test whether they improve downstream results.

How would you evaluate a model that extracts business insights from conversations?

What to coverThe posting emphasizes evaluation strategies and turning conversations into searchable, actionable data. Suggested preparation: distinguish model-level quality from usefulness to sales or service teams, and explain how you would inspect errors before choosing improvements.

How would you design real-time audio understanding for natural speech interaction?

What to coverThe role explicitly includes natural speech interaction and real-time audio understanding. Suggested preparation: describe how you would balance responsiveness, accuracy, and compute demands, and how you would measure those trade-offs.

What would you consider when building an agent that operates on real-world audio?

What to coverBuilding agents capable of operating natively on audio inputs is a stated responsibility. Suggested preparation: explain how you would assess whether the agent understood an input, handle uncertain outputs, and evaluate its behavior on representative audio.

How would you build a scalable pipeline for audio model training and evaluation?

What to coverThe posting asks for scalable data pipelines supporting both training and evaluation. Suggested preparation: cover reliable ingestion, reproducible preprocessing, dataset versioning as a possible approach, and how you would detect pipeline failures or inconsistent results.

How would you structure Python-based model development and deployment?

What to coverStrong Python skills, production deployment, and hands-on experience with frameworks such as PyTorch or TensorFlow are listed. Suggested preparation: describe how you would keep training and inference behavior consistent, test changes, and make a model maintainable in production.

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

What to coverProduction monitoring and maintenance are part of the stated lifecycle. Suggested preparation: walk through checks on incoming data, inference behavior, model outputs, and recent changes before deciding whether to fix the pipeline, adjust the model, or revisit evaluation.

How would you decide where LLMs fit within an audio ML application?

What to coverThe role includes collaboration on LLM and audio ML applications, but the posting does not prescribe an architecture. Suggested preparation: explain when an LLM component might add value, how you would evaluate its outputs, and what latency or reliability costs could change the design.

How would you improve an ML system's scalability without losing performance visibility?

What to coverThe posting calls for improving ML infrastructure, tooling, scalability, and performance. Suggested preparation: identify the bottleneck you would measure first, compare possible changes, and explain how monitoring would show whether the change helped.

Behavioural practice

Tell me about a time you translated a customer's need into an ML solution.

What to coverDirect customer engagement to identify needs and deliver solutions is part of the role. Suggested preparation: use a real situation, your questions and decisions, and the resulting outcome; distinguish what the customer requested from what you chose to build.

Describe a time customer feedback changed your technical approach.

What to coverThe posting explicitly calls for gathering customer feedback. Suggested preparation: give the situation, the feedback you received, the action you took, and the result, including how you determined whether the change addressed the need.

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

What to coverCross-functional collaboration on the AI stack and tooling is a stated responsibility. Suggested preparation: describe the situation, how you resolved differing priorities or assumptions, your contribution, and the outcome.

Describe a time you took ownership of a model after deployment.

What to coverThe role expects end-to-end ownership through production maintenance and monitoring. Suggested preparation: use a real situation to explain what you observed, what you did, how you coordinated any response, and what changed as a result.

Tell me about a time you improved an ML workflow for scale or performance.

What to coverContinuous improvement of ML infrastructure and processes is part of the work. Suggested preparation: state the original constraint, the action you took, the trade-offs you considered, and the measured or observed result.

Sharpen your knowledge

Topics to prepare

Production ML lifecycle and architecture
Map the posted lifecycle—data acquisition, preprocessing, training, inference, deployment, and monitoring—onto one system diagram. Identify where failures could affect searchable conversation insights; the role explicitly owns this lifecycle.
Audio and unstructured-data modeling
Compare two approaches to extracting useful information from recorded conversations, noting how recording quality and preprocessing could affect results. Audio-focused ML and unstructured voice data are central to the posting.
Real-time speech systems and audio agents
Sketch how audio input could reach a model and produce a response, then identify latency and failure points. This prepares you to discuss the stated work on natural speech interaction, real-time understanding, and audio-input agents.
Evaluation and feature engineering
Draft an evaluation plan for a conversation-insight model: define what a useful output means, how examples would be labeled, and what errors to inspect. The posting explicitly calls for evaluation strategies and feature engineering.
Scalable pipelines and monitoring
Draw a training-and-evaluation pipeline and explain what you would monitor after deployment, including data quality and model behavior. Scalable pipelines, performance, and production monitoring are stated priorities.

Skills to revise

Python model implementation
In Python, use PyTorch or TensorFlow to build a small model-training and evaluation loop on sample unstructured data, and explain its preprocessing choices. Python and hands-on use of either framework are explicit qualifications; this exercise is a preparation suggestion.
Audio preprocessing and analysis
Practice loading sample audio, preparing model inputs, and examining how noisy or incomplete recordings change outputs. This suggested exercise targets the posted audio-model development and real-world voice-data work.
Training and evaluation pipeline construction
Write a repeatable pipeline that separates data preparation, training, and evaluation, then test how it handles a bad input. The role explicitly involves scalable pipelines and end-to-end ownership.
Inference and deployment walkthrough
Build or outline a callable inference flow for an audio-derived input; rehearse explaining latency tradeoffs, deployment steps, and monitoring signals. This reflects the posting’s production deployment and real-time inference requirements; no particular serving tool is required.
Customer-facing technical communication
Role-play a customer describing an unmet need, then summarize the need, propose a measurable model outcome, and explain a limitation in plain language. Direct customer and stakeholder engagement is an explicit part of the job.

Your preparation checklist

Map the stated qualifications to evidence
The posting asks for a related degree, 1–6 years of professional ML engineering experience, strong Python, and PyTorch or TensorFlow. Prepare a one-page evidence sheet with accurate examples and dates; do not claim experience you cannot substantiate.
Prepare a production ML project walkthrough
The role emphasizes end-to-end ownership. Select one relevant project and explain your decisions from data acquisition and preprocessing through training, deployment, inference, monitoring, and maintenance, including what you personally owned.
Work through a real-time audio system design
The role covers natural speech interaction, audio understanding, and agents operating on real-world audio. As interview practice, sketch an architecture for turning field conversations into searchable insights, with latency, reliability, and deployment tradeoffs.
Document an evaluation and pipeline example
The posting calls for scalable training and evaluation pipelines, feature engineering, and model evaluation. Prepare a diagram or brief write-up showing data flow, evaluation criteria, and how you would detect performance changes in production.
Choose concrete implementation examples
Prepare to discuss Python and a framework such as PyTorch or TensorFlow in code-level terms. If applicable, add examples involving cloud infrastructure or real-time inference; the named tools, including AWS, TypeScript, and FastAPI, are described as preferences or pluses, not universal requirements.
Rehearse a customer-facing explanation
The role involves working directly with customers to identify needs and gather feedback. Prepare a short example of how you would clarify a customer problem, explain an ML limitation plainly, and turn feedback into a testable improvement.
Bring questions about the work
Ask which audio inputs and use cases matter most, how the team evaluates insight quality and real-time performance, and where this engineer would own the model lifecycle. These are questions to resolve with the employer, not assumptions about its processes.

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…