Uber Senior Software Engineer Interview Experience — AI/LLM Onsite With an LLM Pipeline Retry Design Question

Just wrapped up an onsite loop for a Senior SWE role focusing on AI/LLM engineering. Thought I'd share a quick breakdown of the experience and one of the core technical questions that came up, since the format was a bit different from standard LeetCode grinding.

Notable Technical Question

In the coding/system architecture round, they presented a practical scenario rather than a standard algorithm problem:

The Problem: You have a multi-step LLM extraction pipeline where step N depends on structured JSON output from step N-1. Occasionally, the upstream model output deviates slightly from the expected schema, causing downstream failures. How would you architect a validation and retry mechanism that minimizes token waste and processing latency without blowing past strict timeouts?

Key Discussion Points:

1. Schema Enforcement: Implementing strict runtime validation (e.g., using Zod or JSON schemas) right at the boundary of each step.
2. Targeted Self-Correction: Fast-failing malformed responses and using localized correction prompts (sending only the invalid JSON + error schema back to the LLM) rather than re-running the entire upstream context.
3. Caching & Fallbacks: Caching intermediate deterministic transformations to keep overall latency down during retries.

Overall, it felt much more like a real-world engineering review than a whiteboard puzzle.

Anyone else seeing a shift away from traditional algorithm questions toward pipeline resilience and system design in AI roles?