New opportunity

Machine Learning Engineer

WireScreen · New York, United States

About this role

WireScreen seeks a Machine Learning Engineer to develop models supporting entity resolution and a knowledge graph for its global supply-chain and China-related risk intelligence platform. Reporting to the VP of Engineering, the engineer will work with data, enrichment, product and research teams to identify connections among people and organizations and establish a reliable source of truth.

Responsibilities

Refine entity-resolution algorithms and expand the knowledge graph using alternative data to reveal relationships within the Chinese economy.
Train, test and deploy models processing tens of millions of records daily; contribute to MCP servers and agentic pipelines that classify, cluster and enrich millions of entities.
Partner with Product to build evaluation harnesses for classical ML and agentic systems. Integrate agent workflows into internal tools to improve research speed and scale, and help scale data ingestion and platform analytics.

Required qualifications

4+ years working on clustering-type ML problems, ideally knowledge graphs or entity resolution; related domains may include recommendation engines, cohort analysis, or outlier and anomaly detection.
End-to-end experience taking ML services from experimentation, training, testing and tuning through production deployment and ongoing operation. Relevant model families include clustering, classification, regression, dimensionality reduction, embeddings, similarity methods, ensembles, NLP and deep learning.
Significant Python and SQL experience.

Preferred qualifications

Experience with frontier models or task-specific LLM fine-tuning; large, heterogeneous, unstructured datasets; semantic search; computer vision, especially OCR; or linear optimization.
Familiarity with PySpark, Temporal, FastAPI, Scikit-learn, NumPy, Docker, Terraform or Kubernetes. Early-stage startup or B2B SaaS experience is also preferred.

The full-time role is based in New York, NY, with a hybrid office schedule. Estimated compensation is $150K–$195K annually, with equity; actual compensation may vary by experience, skills, qualifications and location. Benefits include 100% company-paid employee medical, dental and vision coverage, FSA and HSA options, a 401(k), paid time off, company-wide holidays and pre-tax commuter benefits.

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

Skills for this role

Machine LearningEntity ResolutionKnowledge GraphsClusteringClassificationRegressionModel EvaluationAgentic SystemsMCPPythonSQLEmbeddingsDimensionality ReductionNearest-Neighbor MethodsKNNSVMNatural Language ProcessingDeep LearningLLM Fine-TuningSemantic SearchComputer VisionOCRLinear OptimizationPySparkTemporalFastAPIScikit-learnNumPyDockerTerraform 4 years of experience working on clustering-type ML problemsideally

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

Improve entity resolution
WireScreen explicitly calls for fine-tuning existing algorithms to identify connections between people and organizations across China. The role centers on clustering-type ML, ideally in entity resolution or knowledge graphs.
Expand the knowledge graph
Use alternative data to map relationships and help establish a reliable source of truth, working across data, enrichment, engineering and product.
Operate ML at daily production scale
Train, test and deploy models processing tens of millions of records daily. The posting also requires experience carrying an ML service from experimentation and tuning through deployment and ongoing operation.
Define useful evaluation with Product
Partner with Product to implement evaluation harnesses for both classical ML and agentic systems, so model and workflow changes can be assessed against their intended use.
Make agent workflows useful to Research
Contribute to existing MCP servers and agentic pipelines that classify, cluster and enrich millions of entities; integrate agent workflows into internal tools to improve Research’s speed and scale.

Likely challenges to think through

Balancing matches against false connections
Preparation hypothesis: resolving people and organizations across alternative data may force a trade-off between finding hidden links and adding incorrect graph relationships. Prepare to explain how you would review ambiguous matches, assess error types and decide when evidence is sufficient.
Keeping models reliable at volume
Preparation hypothesis: processing tens of millions of records daily may make thorough testing and iteration harder to balance with operational cost and throughput. Prepare an end-to-end example covering deployment, monitoring, failures and subsequent tuning, without assuming WireScreen’s specific infrastructure.
Measuring agent speed without losing trust
Preparation hypothesis: faster Research workflows may not be useful if classification or enrichment errors go unnoticed. Prepare to discuss how an evaluation harness could measure both task quality and time saved, with human review where uncertainty matters.

Practice your answers

Technical practice

How would you refine entity resolution for people and organizations across inconsistent data sources?

What to coverSuggested preparation: Explain candidate generation, matching signals, and how you would balance missed connections against false merges. Discuss how you would inspect errors and tune an existing algorithm rather than assume a particular model is in use.

How would you add alternative data to a knowledge graph without weakening its reliability?

What to coverSuggested preparation: Address entity and relationship definitions, conflicting evidence, and how to retain the source of a proposed connection. Explain how you would decide whether new data improves the graph’s usefulness as a source of truth.

How would you choose and assess a clustering approach for millions of entities?

What to coverSuggested preparation: Compare relevant clustering or similarity approaches in terms of record quality, scale, and the kinds of errors they produce. Explain how you would assess cluster quality when complete labels are unavailable.

How would you design a model workflow that processes tens of millions of records daily?

What to coverSuggested preparation: Walk through training, testing, deployment, and ongoing operation at the volume stated in the posting. Discuss throughput, data freshness, and what you would measure to detect a failing or degraded run.

What would you include in an evaluation harness for a classical entity-resolution model?

What to coverSuggested preparation: Describe representative test data, useful matching metrics, and checks for performance across different record types. Explain how the harness would help Product compare model changes against the cost of false and missed connections.

How would you evaluate an agentic system that classifies or enriches entities?

What to coverSuggested preparation: Propose checks for output quality, consistency, and cases that need human review. Contrast these checks with classical model evaluation, and explain how you would detect regressions before relying on an agent workflow.

How would you integrate an agent workflow into an internal research tool?

What to coverSuggested preparation: Explain how an agent and the existing MCP servers could support classification or enrichment while making results inspectable by researchers. Discuss trade-offs between research speed, reliability, and opportunities for human review.

How would you investigate a production model that starts creating incorrect entity connections?

What to coverSuggested preparation: Trace possible changes in input data, model behavior, and downstream graph updates. Explain how you would contain the problem, identify affected records, and verify a correction.

How would you use Python and SQL to investigate a data-ingestion or model-quality issue?

What to coverSuggested preparation: Describe how you would query affected records and use Python to inspect patterns, reproduce the issue, and test a fix. Keep the example focused on the posting’s data ingestion, entity resolution, or platform analytics work.

When might semantic search or OCR help with messy source data, and what would you validate?

What to coverSuggested preparation: Treat these as optional areas named in the posting, not required parts of the system. Discuss how extracted or retrieved information might supply matching evidence, and how you would check errors before using it to add graph relationships.

Behavioural practice

Tell me about a time you aligned product and engineering on how to evaluate an ML system.

What to coverSuggested preparation: Describe the situation, your actions to agree on success measures and trade-offs, and the result. Focus on a real example rather than assuming WireScreen uses a particular evaluation process.

Tell me about a time different data sources disagreed about an entity or relationship.

What to coverSuggested preparation: Set out the situation, how you worked with relevant partners to resolve or represent the uncertainty, and the result. If your example comes from another domain, explain its relevance without claiming it was knowledge-graph work.

Tell me about a time you made a research or analytical workflow faster without losing trust in its output.

What to coverSuggested preparation: Explain the situation, the workflow change you made, and the result for its users. Include how you checked quality or preserved a way to review questionable outputs.

Tell me about a production ML problem you had to diagnose after deployment.

What to coverSuggested preparation: Describe the situation, your investigation and response, and the result. Be clear about your own actions and what the incident changed in ongoing operation.

Tell me about a time you prioritized ML work amid changing requirements.

What to coverSuggested preparation: Use a real situation to explain the competing needs, the decision you made with stakeholders, and the result. This is a suggested practice topic based on the posting’s early-stage setting and cross-team scope, not a stated interview question.

Sharpen your knowledge

Topics to prepare

Entity resolution decisions
Posting: Refine entity-resolution algorithms to connect people and organizations. Suggested revision exercise: Given ambiguous names and conflicting attributes in a hypothetical dataset, explain how you would decide which records represent the same entity and investigate false matches.
Knowledge graph reliability
Posting: Expand a knowledge graph with alternative data and establish a source of truth. Suggested revision exercise: Sketch how a new relationship would be linked to its supporting records, and explain how you would handle contradictory evidence.
Clustering and similarity methods
Posting: Clustering-type ML experience is required; embeddings, dimensionality reduction and nearest-neighbor methods are among the relevant model families. Suggested revision exercise: Compare two approaches to grouping messy organization records, including when each might incorrectly merge or split entities.
Evaluation of models and agents
Posting: Work with Product to define evaluation harnesses for classical ML and agentic systems. Suggested revision exercise: Outline a labeled test set and separate checks for entity-match quality, agent classification errors and changes between versions.
Production ML at record scale
Posting: Train, test and deploy models operating on tens of millions of records daily, with responsibility extending beyond deployment. Suggested revision exercise: Walk through an experiment-to-operation plan and identify what you would measure if daily processing or output quality degraded.
Agent workflows and MCP context
Posting: Contribute to existing MCP servers and agentic pipelines that classify, cluster and enrich entities, and bring agent workflows into internal research tools. Suggested revision exercise: Diagram a hypothetical enrichment workflow, marking where an agent obtains data, produces a result and needs a quality check.

Skills to revise

Python record-processing practice
Posting: Significant Python experience is required. Suggested exercise: Write a small program that normalizes hypothetical person and organization records, proposes possible matches and reports cases needing review.
SQL investigation
Posting: Significant SQL experience is required, and the role spans data systems and ingestion. Suggested exercise: Query a sample records-and-relationships dataset to find duplicate candidates, missing identifiers and relationships supported by conflicting records.
Model experiment and error analysis
Posting: The role requires end-to-end ML experience, including experimentation, training, testing and tuning. Suggested exercise: Train and compare two simple clustering or similarity approaches on a sample entity dataset; inspect erroneous merges and splits before selecting a version.
Evaluation harness implementation
Posting: Implement evaluation harnesses with Product for classical ML and agentic systems. Suggested exercise: Build a repeatable test that scores labeled entity matches and records agent classification outcomes, then compare two runs and summarize regressions.
Deployment and operational checks
Posting: Deploy ML models and operate them beyond launch at substantial daily record volume. Suggested exercise: Package a small batch inference job, test it on increasing input sizes and specify checks for throughput, failed records and changes in match quality.

Your preparation checklist

Prepare a clustering project walkthrough
Choose a real project, if available, and explain the data, method, baselines, error analysis and measured results. WireScreen asks for 4+ years on clustering-type problems, ideally entity resolution or knowledge graphs.
Practice an entity-resolution design exercise
Sketch how you would identify when records refer to the same person or organization, handle ambiguous matches and incorporate alternative data into a knowledge graph. These are preparation choices for the role, not a stated interview exercise.
Document an end-to-end production ML service
Prepare a diagram or walkthrough covering experimentation, training, testing, tuning, deployment and ongoing operation. Explain how you measured reliability and handled scale; the role includes models processing tens of millions of records daily.
Gather concrete Python and SQL evidence
Select examples you can explain in detail, such as Python model or data-processing code and SQL used to inspect or transform records. Both are stated requirements; include performance or correctness results where available.
Draft an evaluation-harness proposal
Outline datasets, baselines, metrics and failure-case review for entity resolution and an agentic classification or enrichment workflow. WireScreen says this work is done with Product, so identify decisions that would need Product input.
Map an agent workflow into a research tool
As a practice exercise, draw a small workflow that classifies or enriches entities, shows where an MCP server could fit and flags uncertain outputs for researcher review. The posting mentions existing MCP servers and internal tools, but does not specify their architecture.
Write targeted questions for the team
Ask how the data, enrichment, Product and Research teams establish a source of truth; what evaluation gaps matter most; and where ingestion or researcher workflows currently limit scale. These questions connect directly to the advertised responsibilities.

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…