ML Systems Codebase Deep Dive

A no-coding HM technical screen for an ML Systems Engineer role that turns into a deep review of one recent project and its codebase. The interviewer probes architecture, deployment, data collection, cross-evaluation, database choices, agentic-system rationale, FastAPI / LangGraph state handling, PDF embedding, similarity search, and paper retrieval. Requirements Prepare to walk through one recent ML / agentic-systems project as if the interviewer is reviewing the implementation with you. The round may stay on a single project for the full screen. Be ready to cover: Overall architecture and deployment path. How data is collected, cleaned, evaluated, and compared across alternatives. What database or storage layer the project uses, and why that choice fits the workload. Why an agentic architecture was chosen instead of a simpler pipeline. Backend framework details, including FastAPI-style API boundaries and how state is carried through a LangGraph-style workflow. Retrieval details: how PDFs are embedded, how similarity is computed, and how the system selects useful papers or documents. Failures and trade-offs: what broke, what you changed, and what you would redesign with more time. Notes This is not a generic ML trivia screen. The differentiator is whether you can defend your own implementation choices at the level of codebase structure, data flow, persistence, evaluation, and production constraints. Keep answers concise enough for a 30-minute HM screen. Start with a one-minute architecture map, then let follow-ups go deeper. For every choice, state the alternative you rejected and the constraint that made your choice reasonable. A strong answer names concrete interfaces and boundaries without dumping every file or class. For example: API endpoint -> orchestration graph -> retrieval step -> evaluator -> persistence. If the project uses agents, explain what state is stored, which transitions are deterministic, where model calls happen, and how the system recovers from bad retrieval or tool output. Preparation Pick one recent project and write a compact architecture map: request path, data path, storage, evaluation, deployment, monitoring, and failure modes. Prepare a decision table for the main technical choices: agentic workflow vs fixed pipeline, vector retrieval vs keyword search, database choice, evaluation metric, and deployment shape. Rehearse a codebase tour that can be delivered without screen sharing: core modules, key interfaces, state objects, and where the hardest bug lived. For retrieval-heavy projects, be ready to explain embedding generation, chunking, similarity scoring, candidate selection, and cross-evaluation.