This is what you get after every interview.
Not a single averaged score — a per-concept depth map with the evidence behind every rating. Switch between a technical and a finance interview below; both candidates are fictional, but the format, depth scale, and evidence trail are exactly what your reports contain.
Senior AI Engineer
Summary: Strong senior signal on the concepts this role depends on. Retrieval-augmented generation and prompt engineering both reached Expert under ceiling probing. One concrete gap — evaluation methodology, which stalled at Working — is coachable and worth a targeted follow-up question, not a rejection.
Retrieval-augmented generation
Held at the ceiling under adversarial probing. Reasoned correctly about chunking strategy, retrieval failure modes, and when reranking earns its latency across three escalating scenarios; diagnosed the hallucination-from-missing-context case in the final prompt without hints.
“…the model isn’t wrong, the retriever is — it never surfaced the clause, so grounding can’t help. I’d fix recall first with a hybrid BM25 + dense retrieve, then rerank, before touching the prompt.”
Prompt engineering
Strong. Argued context-window budgeting with real consequences, structured a tool-use prompt that degrades gracefully on malformed output, and pushed back correctly on few-shot examples that were leaking the answer format.
“…don’t few-shot the exact output — you’re teaching it to pattern-match the examples. Give it the schema and one contrasting case, and constrain with a stop sequence instead.”
Evaluation & testing
Gap. Comfortable running an offline eval set, but conflated retrieval recall with answer faithfulness and could not describe how to catch a regression that only shows on out-of-distribution inputs. Stopped advancing past aggregate accuracy.
“…if the eval score is high we’re good, right?…” (incorrect — high retrieval recall said nothing about faithfulness; a partial self-correction followed)
Vector search & embeddings
Reasons about trade-offs comfortably. Correct on ANN index choice (HNSW vs IVF), the recall/latency trade, and embedding-model drift on re-indexing; slowed on a metadata-filtered search edge case.
“…cosine on normalized embeddings, HNSW for the recall we need; but re-embed with a new model and you have to re-index everything — the vectors aren’t comparable across models.”
Inference: latency & cost
Explained token streaming, prompt-prefix caching, and the batching-vs-tail-latency trade-off correctly; the interview time-boxed before the final probe, so confidence is moderate — depth may be higher than measured.
“…stream tokens so time-to-first-token stays low, and cache the system-prompt prefix; batching lifts throughput but you pay it back in tail latency.”
Every quote above links to its timestamp in the recorded session in a real report — the rating is an argument you can check, not a number you have to trust.
How to read the depth scale
Four steps, in plain words — each one a higher bar than the last. The same scale applies to every technology and every professional domain.
Knows the terms and can define them.
Applies it correctly in everyday work.
Explains why, weighs trade-offs, catches mistakes.
Holds up under the hardest follow-ups.