AI Engineer

What matters most

• Support investment evaluation

  The role develops AI/ML systems to identify and evaluate biomedical investment opportunities, with systems also informing decisions in medical and climate technologies. Prepare to explain how a model’s output could support, rather than replace, an investment judgment.

• Deliver end-to-end applications

  The posting calls for AI applications from concept through deployment, including frontend and backend components. Prepare one concise account of design choices across the application, model, deployment and monitoring lifecycle; AWS AI platforms, ECS and Kubernetes are named requirements.

• Model diverse data and prediction tasks

  Responsibilities span time-series, business and image data, deep learning optimization and multiple prediction tasks. Prepare to explain how you would choose a baseline, evaluate performance and investigate data or model errors for different data types.

• Build with proprietary research and agentic workflows

  The role includes building and enhancing large language models using proprietary research data. It also requires experience with agentic coding and advanced AI workflows, MCPs and agent swarms. Prepare to distinguish what each approach accomplishes and how its results would be checked.

• Collaborate across disciplines

  The posting emphasizes communication, investigative problem-solving and a mentorship-driven environment; the company describes collaboration with investors and scientific and technical specialists. Prepare to explain technical findings, limitations and next investigative steps in terms a non-ML colleague can use.

Likely challenges to think through

• Combining unlike data without misleading signals

  Likely preparation hypothesis: time-series, business, image and proprietary research data may differ in quality, structure and coverage. The trade-off is gaining richer predictions without introducing unreliable comparisons or leakage. Prepare to discuss data checks, task-appropriate baselines and evaluation that reflects the intended decision.

• Balancing latency and model complexity

  Likely preparation hypothesis: prediction systems operating from milliseconds to daily timescales will not all tolerate the same model or infrastructure cost. Prepare to explain when a simpler or precomputed prediction is preferable to a more complex model, and how you would measure latency alongside predictive performance.

• Keeping AI workflows useful and verifiable

  Likely preparation hypothesis: LLMs and agent-based workflows built around proprietary research could produce plausible but weakly supported outputs. The trade-off is workflow speed versus dependable evidence for investment analysis. Prepare to describe source-grounded evaluation, failure handling and how you would communicate uncertainty.

Topics to prepare

• Machine learning fundamentals

  The posting explicitly calls for regression, classification, and clustering. Revise when each fits an investment-research question; practice explaining suitable evaluation measures and failure modes using a hypothetical biomedical dataset.

• Deep learning across data types

  The role covers time-series, business, and image data, with neural networks and PyTorch required. Compare plausible model approaches for each data type and explain how you would diagnose poor predictive performance.

• Language models and proprietary research

  The role involves building and enhancing large language models trained on proprietary research data. As preparation, sketch how you would separate training, evaluation, and access to research data, and explain how you would check whether outputs support reliable analysis.

• AI/ML development lifecycle

  Full lifecycle experience and applications from concept through deployment are explicit requirements. Trace a hypothetical prediction task from data preparation and model selection through evaluation, deployment, and performance monitoring; monitoring is a preparation suggestion, not a stated requirement.

• Prediction latency and data structures

  The posting specifies prediction timescales from milliseconds to daily and a strong foundation in algorithms and data structures. Work through the trade-offs between these timescales, then identify where data access, computation, or model size could limit response time.

• Advanced AI workflow design

  Professional experience with AI workflows, MCPs, and agent swarms is requested. Diagram a hypothetical multi-agent research workflow and explain its handoffs, failure points, and checks; the diagram is a revision exercise, not a prescribed company architecture.

Skills to revise

• Implement and evaluate a PyTorch model

  Using synthetic time-series or image data, write a small Python training and evaluation loop, inspect errors, and propose one performance improvement. This practices the posting’s Python, PyTorch, and prediction-model requirements without assuming access to proprietary data.

• Build a full-stack prediction prototype

  As a practice exercise, connect a JavaScript frontend to a Java backend that calls a Python prediction component, then explain the request and error paths. The posting calls for full-stack AI applications and proficiency in all three languages; this particular division of work is only a suggested exercise.

• Plan an AWS deployment

  Draw a deployment plan showing where SageMaker, Bedrock, ECS, and Kubernetes could fit, and explain which components you would actually need for a sample AI application. These named platforms and orchestration technologies appear in the requirements; the plan need not claim they are all used together at RA Capital.

• Test a research-data language model

  With invented research notes, formulate example investment-analysis questions and a rubric for checking answers against those notes. This practices the stated language-model and proprietary-research work; the test notes are a substitute, not a claim about company data.

• Practice agentic coding

  Use Claude Code or Codex on a small, nonproprietary coding task, review its changes, and explain how an MCP or agent handoff would be tested. The exercise targets the posting’s explicit agentic-coding and advanced-workflow requirements.

• Investigate and explain model failures

  Benchmark a sample prediction pipeline, identify a bottleneck or erroneous output, and summarize the evidence and proposed fix for a nonengineering collaborator. This combines the role’s performance-optimization work with its stated investigative and English communication requirements.

• Build a requirements evidence sheet

  The posting specifies 2–3 years of software engineering, AI-focused education or equivalent experience, Python, JavaScript, Java, PyTorch, AWS AI platforms, and deployment tools. Prepare a one-page map from each requirement to a specific example you can discuss; mark any gaps rather than inventing evidence.

• Prepare an end-to-end application walkthrough

  Select a real project, if available, and diagram its frontend, backend, model, data flow, deployment, and your contribution. The role explicitly covers AI applications from concept through deployment.

• Practice a predictive-model design exercise

  Using a hypothetical biomedical investment question, outline how you would handle time-series, business, or image data. Prepare a baseline, a PyTorch approach where appropriate, evaluation metrics, and checks for missing data and leakage; these are preparation choices, not a stated interview exercise.

• Outline a research-data LLM approach

  The role calls for enhancing LLMs with proprietary research data. Using only public or synthetic examples, prepare a diagram of data handling, model or retrieval choices, evaluation, and how you would check unsupported outputs.

• Sketch deployment across prediction timescales

  Prepare an architecture and latency budget for a hypothetical prediction service operating from milliseconds to daily runs. Explain where SageMaker, Bedrock, ECS, or Kubernetes might fit, and the trade-offs; the posting names these technologies but does not prescribe one architecture.

• Document an agentic-workflow example

  The posting asks for professional experience with agentic coding, advanced AI workflows, MCPs, and agent swarms. Prepare a concrete account of work you have actually done, including tool use, testing, oversight, and limitations; distinguish experience from proposed designs.

• Rehearse a cross-disciplinary explanation

  Prepare a two-minute explanation of a technical decision and its uncertainty for non-engineering collaborators. The role connects AI systems to investment decisions and describes collaboration with investors, physicians, biologists, and engineers.

• Prepare role-specific questions and logistics

  Write questions about research-data access, success measures for prediction systems, and collaboration and mentorship. Also be ready to discuss the stated Boston hybrid arrangement and work-authorization requirement without assuming an exception or assessing your eligibility.