AI Engineer Interview Guide 2026: Build These 5 Apps
An AI engineer interview guide built around five portfolio apps, the questions each one answers, and why the RAG app is the one candidates get wrong.
Proving it works and keeping it working: eval, tracing, security, routing.
8 posts
An AI engineer interview guide built around five portfolio apps, the questions each one answers, and why the RAG app is the one candidates get wrong.
Batch re-indexing leaves your RAG index lagging the source, so answers go stale. Stream changes with CDC, upsert by hash, and pin versions to stay fresh.
Model routing cuts LLM cost and adds failover — but each request can hit a different model. Pin the context and log the route, or fallbacks hide regressions.
In RAG, retrieved documents are untrusted input. Indirect prompt injection turns your knowledge base into an attack surface — defend at ingestion and retrieval.
You can't fix a RAG regression you can't see. Instrument every request as one trace, watch the right signals, and run online evals to catch it fast.
Span trees and token counts tell you what your RAG pipeline did. Reproducing the incident needs the corpus state it ran against. Instrument both.
You can't tell if a retrieval tweak improved RAG by reading a few answers. Measure it in two metric layers, then gate the ship on a paired test.
KV and prefix caching speed up prefill by reusing exact token prefixes. They do not make your RAG answers reproducible — that is a separate layer, upstream.
Inherent on Substack
Everything you need to know about AI, delivered to your inbox. Every week.