Machine Learning Engineer II

What matters most

• Build end-to-end RAG pipelines

  The role designs pipelines integrating proprietary chunking algorithms, embeddings, vector databases, and retrieval agents. Prepare to explain the design choices and interfaces across those stages.

• Improve retrieval over large datasets

  DRIVE builds retrieval systems over large-scale proprietary, unstructured data. The posting specifically emphasizes embedding techniques, similarity search, vector indexing, and retrieval evaluation.

• Orchestrate context-aware LLM responses

  The role combines retrieval, generation, and ranking for agentic search and experiences such as question answering and deep research. Prepare to discuss how each stage affects response quality.

• Ship and maintain production ML systems

  This is a hands-on, full-lifecycle role. The posting calls for experience with data processing, training, inference, maintenance, evaluation, versioning, and experimentation, alongside ML Operations collaboration on automation.

• Connect ML work to user needs

  The engineer works with Product and Design to improve user experiences and meet business objectives. Prepare to explain technical decisions in terms of their effect on a search or research workflow.

Likely challenges to think through

• Balancing retrieval quality with scale

  Likely preparation challenge, not a confirmed team problem: chunk size, embedding choice, and index strategy can improve relevance while increasing storage, latency, or update complexity. Prepare a concrete comparison of alternatives and how you would evaluate them on representative unstructured data.

• Diagnosing errors across a RAG pipeline

  Likely preparation challenge: a poor answer may stem from retrieval, ranking, or generation, making end-to-end scores alone hard to act on. Prepare a staged evaluation approach that isolates retrieval failures from response-quality failures.

• Turning experiments into maintainable systems

  Likely preparation challenge: iterating on models and retrieval methods can conflict with reliable, reproducible production operation. Prepare to explain how versioning, evaluation, documentation, and automation would support changes, and where Product, Design, and ML Operations input would matter.

Topics to prepare

• End-to-end RAG architecture

  The DRIVE role integrates chunking, embeddings, vector databases, retrieval agents, ranking, and generation. As a revision exercise, sketch the flow from an unstructured document to a cited answer and explain where retrieval or generation can fail.

• Chunking and embedding trade-offs

  The posting emphasizes proprietary chunking algorithms, advanced embeddings, and large-scale text retrieval. Compare two chunking approaches on a sample document set; reason about lost context, retrieval precision, and indexing cost. This is a preparation exercise, not a stated employer task.

• Similarity search and vector indexing

  Experience with vector databases, similarity search, and indexing algorithms is explicitly requested. Revise how index and search choices affect recall, latency, and scale; work through a hypothetical large-dataset retrieval bottleneck.

• Retrieval and agent evaluation

  The role investigates unstructured-data retrieval evaluation and lists AI agent evaluation among favored techniques. Design a small set of questions with relevant documents, then explain how you would measure retrieval quality separately from answer quality.

• Agentic search and GraphRAG

  The team builds agentic search pipelines and investigates GraphRAG. Trace when a retrieval agent might need another search step, and compare a graph-based retrieval path with vector search for a question spanning related documents.

• Production ML lifecycle

  The posting requires experience across data processing, training, inference, maintenance, evaluation, versioning, and experimentation. Draw a lifecycle diagram and identify what must be monitored or updated when source documents or embedding models change.

Skills to revise

• Build a Python text-processing pipeline

  Strong Python skills and working knowledge of PyTorch, Transformers, and HuggingFace are requested. As practice, write a small pipeline that cleans and chunks documents, generates embeddings with a relevant model, and records inputs and model version.

• Implement and inspect vector retrieval

  The posting names PostgreSQL/PGVector, OpenSearch, and Pinecone as vector-database examples. Using one of these, practice indexing sample embeddings, running similarity queries, and inspecting missed relevant results; no particular database is mandated.

• Orchestrate retrieval, ranking, and generation

  Experience with LLM orchestration frameworks such as LangChain or LlamaIndex is requested. Practice wiring a retriever to a ranking step and response generation, then inspect which retrieved passages influenced the answer.

• Run a reproducible retrieval experiment

  Evaluation, versioning, and experimentation are explicit pipeline requirements. Compare two chunking or indexing configurations on the same question set, record retrieval results and configuration versions, and explain the quality–latency trade-off.

• Plan an operations handoff

  The role involves maintaining production systems and collaborating with ML Operations on lifecycle automation. Draft a brief handoff for an embedding-index update covering validation, deployment, monitoring, and rollback; treat the document as interview preparation, not a specified Kensho process.

• Select a production ML project walkthrough

  The posting calls for experience designing, shipping, and maintaining ML and retrieval systems. Prepare a concise walkthrough of one real project, covering your contribution, scale, tradeoffs, results, and maintenance lessons; use a different example if you have not shipped RAG.

• Sketch an end-to-end RAG design

  As practice for the DRIVE team’s work, draw a proposed flow from unstructured documents through chunking, embeddings, indexing, retrieval agents, ranking, and response generation. Be ready to explain how components interact and where you would expect bottlenecks.

• Prepare a retrieval tradeoff comparison

  The role investigates vector search, chunking, indexing, and GraphRAG. Create a small comparison of plausible approaches, stating when you would test each, how you would measure retrieval quality and latency, and what evidence would change your choice.

• Outline an evaluation plan

  Draft a proposed test set and measures for retrieval relevance and context-aware answers, including failure examples. Explain how you would compare pipeline versions and assess whether an agentic search change improves results.

• Gather implementation evidence

  Choose specific examples that show your Python work, relevant ML or LLM orchestration frameworks, and similarity search or vector indexing experience. For each, note what you built yourself, how you tested it, and any limits to your experience.

• Map the ML system lifecycle

  The posting includes data processing, training, inference, evaluation, versioning, experimentation, and maintenance. Prepare a diagram or brief account showing how you handled those stages and where automation and ML Operations collaboration fit.

• Draft collaboration stories and employer questions

  Prepare one example of working with product or design partners on a user-facing tradeoff. Ask how DRIVE defines search success, evaluates unstructured-data retrieval, and divides lifecycle responsibilities between engineers and ML Operations.