Defend

Semantic search and RAG, +35% accuracy

Older work, but it tests whether you can compare models and vector stores with reasons.

On your resume

Built a production semantic-search and RAG platform that improved search accuracy by 35% while reducing response latency and manual discovery effort.

Cross-questions

3
A strong answer covers
  • Cost, latency and quality per task, measured on your own eval set.
  • Vendor resilience and routing.
Avoid

“We tried both and picked the better one,” with no criteria.

A strong answer covers
  • Managed and fast to start, with metadata filtering and namespaces.
  • Trade-offs: cost at scale, data residency and lock-in.
Avoid

Not knowing the alternatives.

A strong answer covers
  • Blocking to cut candidate pairs, embedding similarity to find candidates, and an LLM to adjudicate ambiguous pairs.
  • Precision and recall measured on labeled pairs, with cost under control.
Avoid

Sending every pair to an LLM.

Your numbers

Interviewers probe numbers first. Write exact values so you never have to improvise.

Next claim