Project Architecture Deep Dive (AI Engineer / Backend)

Salesforce dedicates a round to one past project and drills into architecture, trade-offs, challenges, scale, and hindsight. AI Engineer versions favor agent orchestration and ML-system decisions; a backend version uses a notification system to probe service decomposition, event-driven messaging, failure handling, consumer concurrency, and production operations. Requirements Pick one project to drive the conversation. Choose for architectural complexity + visible challenges and trade-offs — not for headline impact alone. Be ready to discuss: System architecture end-to-end. Key technical decisions and the alternatives you rejected (and why). Specific challenges encountered and how you resolved them. Scale numbers (QPS, latency, data volume) if relevant. What you would do differently in hindsight. For AI Engineer specifically, the interviewer is biased toward projects involving Multi-Agent Orchestration, LLM serving, or significant ML system-design components. Strongly recommended to bring or pre-draw an architecture diagram; candidates who walk in with one consistently report a smoother round. For an event-driven backend project, prepare precise answers to questions such as: Why split a monolith into microservices? Why choose asynchronous event-driven communication over synchronous calls? Why use Kafka rather than RabbitMQ or Redis? How does the notification system work end to end? How do you detect, debug, and recover from a poison message? Where would a circuit breaker belong? How can a consumer race condition occur, and how do you prevent it? How would you implement a dead-letter queue for Kafka? What is the on-call response after a pager alert? Where do the outbox and sidecar patterns fit? Notes The round is closer to a system-design discussion than a slide-show — interviewers will redraw or annotate your diagram. Prepare for whiteboard / shared-canvas use. Trade-off questions are the most predictive scoring axis. "Why this DB, why this model, why this concurrency choice" — have a defended answer for each non-trivial decision in the project, even if the answer is "it was the default and we never had to revisit it" with justification. Failure / challenge story is expected. Pick something with real complexity (a debugging saga, a scaling crisis, a model regression) — not "the PR review took longer than expected". Pace: ~10 minutes overview, ~30 minutes deep-dive on whatever the interviewer probes, ~15 minutes on what-you-learned and what-you-would-change. Don't burn 30 minutes on overview. AI Engineer loop specifically values: agent orchestration, tool-use design, evaluation framework, latency / cost trade-offs in LLM serving, RAG architecture. Hook into one of these if your project touches it. Backend interviewers may stay on one resume project across multiple rounds and ask for operational detail: poison-message debugging, dead-letter handling, race conditions, on-call practice, and how architectural patterns fit the actual production system. Preparation Pick 2 candidate projects in advance; rehearse a 10-minute overview for each. Draw the architecture diagram for the primary project — boxes-and-arrows level, with traffic flow and data flow on the same canvas. Memorise it. For each box on the diagram, prepare a 1-sentence trade-off story. Run a mock with a friend at the depth level interviewers expect ("why is this Kafka and not RabbitMQ", "why did you pick fine-tuning over RAG", "why this many shards") — the gap between rehearsed and unrehearsed is large on this round. For an event-driven project, rehearse one concrete failure from detection through mitigation and permanent fix, then defend the broker, retry, idempotency, and concurrency choices.