Applied Data Scientist

Novacore•Remote - Pennsylvania, PA

About The Position

We are looking for an Applied Data Scientist with 3+ years of experience who has shipped models into production systems used by real end users. Insurance experience is not required — if you have built classification pipelines, scoring engines, or document extraction systems in fintech, healthcare, logistics, e-commerce, or any other data-rich, decision-heavy domain, we want to hear from you. The domain knowledge transfer is something we can support; the engineering discipline and deployment experience are what matter most. We are an MGA — a company that designs, underwrites, and distributes specialty insurance products. Think of us as a high-velocity decision engine: thousands of submissions come in, each requiring rapid evaluation against a set of complex, evolving criteria. We are building the AI infrastructure that makes those decisions faster, more consistent, and more accurate. The models you build will run inside that infrastructure.

Requirements

3+ years of experience as a data scientist or ML engineer, with several production deployments where your model ran inside a system used by real end users.
Strong Python skills with production-grade coding practices: modular, tested, version-controlled code — not just notebook-quality work.
Hands-on experience with ML frameworks (scikit-learn, XGBoost, LightGBM, PyTorch, or TensorFlow) and applied knowledge of classification, ranking, regression, and feature engineering for real-world, noisy datasets.
Experience building and maintaining data pipelines that feed production models — scheduled, monitored, and reliable, not just ad hoc EDA scripts.
Familiarity with model deployment patterns: REST APIs (FastAPI or Flask), containerization (Docker), and cloud deployment on AWS, GCP, or Azure.
Proficient in SQL; comfortable pulling and transforming data from a cloud warehouse (Snowflake, BigQuery, or Redshift) as part of feature engineering workflows.
Strong problem-framing instincts: you can take an ambiguous business problem, identify whether ML is the right tool, define the target variable, and scope the modeling approach before writing a line of code.

Nice To Haves

Experience in a domain with high-volume, structured decisioning: fintech (credit scoring, fraud detection, loan underwriting), healthcare (risk stratification, claims adjudication, prior authorization), logistics (routing, demand forecasting), or e-commerce (pricing, recommendation, fulfillment) — these backgrounds translate well and are actively valued.
Prior experience in insurance, insurtech, or an MGA is a genuine plus, but not a requirement.
Familiarity with concepts like loss ratio, submission triage, risk scoring, or policy lifecycle will shorten your ramp time and is worth noting in your application.
Experience with LLMs and document intelligence pipelines: prompt engineering, RAG architectures, fine-tuning, or extraction from semi-structured documents (PDFs, forms, emails).
MLOps tooling experience: MLflow, Weights & Biases, SageMaker, Vertex AI, or equivalent platforms for model tracking, serving, and monitoring.
Knowledge of model explainability techniques (SHAP, LIME) and experience designing systems where model outputs need to be interpretable by non-technical end users.
Experience integrating geospatial, firmographic, or third-party enrichment data as ML features — not just as lookup tables, but as engineered inputs to a model.
Familiarity with algorithmic fairness considerations and regulatory constraints on automated decisioning — relevant in insurance, lending, healthcare, and other regulated verticals.

Responsibilities

Build and maintain the models that evaluate inbound requests in real time — scoring quality, flagging risk signals, and routing items to the right handling path before a human reviews them.
Train classification and ranking models on historical outcome data (accepted, declined, loss events) to predict account quality and prioritize review queues — similar in structure to fraud scoring at fintechs, patient risk stratification in healthcare, or lead scoring in high-volume sales platforms.
Integrate structured and unstructured third-party data signals as model features: geospatial layers, firmographic data, external risk indicators, and document-extracted fields.
Serve model outputs via API so scores and flags appear natively inside workflow tools used by the operations team — your model is a product feature, not a report.
Translate business rules and decisioning criteria into machine-executable logic that can be applied programmatically at intake — moving decisions that currently require human judgment into automated or assisted pathways.
Build and own the feature engineering pipelines that feed these models: normalizing inputs, handling missing data, encoding categorical variables, and enriching records with external data sources.
Develop model explainability layers so end users understand why a record was scored or routed a particular way — a requirement for user trust and, in our industry, regulatory defensibility.
Own the full deployment lifecycle: containerize models, write inference APIs, coordinate with engineering on production integration, and set up monitoring for model drift and performance degradation over time.
Instrument deployed models with logging and alerting so the team detects underperformance before it shows up downstream — the same operational rigor expected in any production ML system, regardless of industry.
Maintain model versioning and a rollback-capable registry so changes can be audited, compared, and reversed — increasingly important in regulated environments.
Contribute to shared ML infrastructure: feature stores, reusable pipeline components, and standardized evaluation frameworks used across the data team.
Build pipelines to extract structured data from unstructured documents: forms, PDFs, emails, and attachments that arrive as part of the intake workflow.
Apply NLP and LLM-based extraction techniques to reduce manual data entry and improve the completeness of records entering the decision workflow — analogous to document intelligence work in healthcare (clinical notes), legal tech (contract extraction), or lending (income verification documents).
Develop confidence scoring for extracted fields so downstream systems know when to auto-populate versus queue for human review.