This is the real report surface InterviewLM generates when a candidate finishes their rounds — the same components our customers see in their dashboard. Every score carries the number of logged events behind it, and every claim quotes the moment in the session that produced it.
Senior · Python · PostgreSQL · AWS · completed 12 July 2026
Maya works the way a strong senior engineer actually works in 2026: she reaches for the AI constantly, but never accepts its output without a reason. Across 47 logged AI interactions she rewrote her own prompt 11 times to add constraints the AI had ignored, and overrode it 4 times — every override was correct. Her weakness is not AI dependence, it is test coverage: she verifies behaviour by reading code and reasoning, not by writing tests that would catch a regression later. Recommend hire at senior level, with testing probed onsite.
Four scored dimensions. “Evidence” is the count of logged session events supporting the score — prompts, edits, terminal commands, test runs — and “confidence” is the evaluator's own certainty given that evidence.
Code Quality
74
Problem Solving
86
AI Collaboration
91
Communication
79
Overall Score
83
out of 100
Each subcategory is scored against the level this specific role requires, so a 74 on a role that needs 65 is a pass and a 58 on a role that needs 70 is a gap.
out of 100
Meets or exceeds the senior bar on 7 of 9 assessed subcategories. The two gaps — automated testing and observability — are coachable and both sit outside the role's critical path.
Automated testing
Backend Engineering
She already finds the right edge cases by reasoning — the gap is converting a found edge case into a committed test. Ask her in the onsite to write the regression test for the race she diagnosed here.
Critical evaluation of AI output
AI Collaboration
Rejected an AI-proposed composite index with a correct cardinality argument, and caught a stale-read race in AI-generated invalidation logic.
Data modelling & SQL
Backend Engineering
Went straight to the generated SQL to locate the N+1 instead of guessing at the ORM layer.
Observations are anchored to what the candidate actually did or said. This is the “evidence, not vibes” part — the reason a hiring manager can disagree with a score and check it themselves.
Diagnosed before coding. Spent the first 6 minutes reading the slow query log and the ORM's emitted SQL, and did not open the editor until she could name the cause.
"Before I touch anything — this endpoint is 1.9 seconds and the query log shows 340 selects. That's an N+1, not a slow query. Let me confirm which relation."
Used the AI adversarially against her own design rather than asking it for an answer.
"Here's my index. Argue against it — what workload makes this the wrong choice?"
Caught an incorrect AI-generated cache-invalidation branch that would have served a stale balance after a concurrent write.
"No — this invalidates after the write returns, so a read in between gets the old balance. That's a race, not a style issue."
Verified by reasoning rather than by test. Having found the race, she fixed it and moved on without adding a test that would catch a regression.
Session shows the race fixed at 00:41 and no test file modified after 00:33.
Abandoned her first approach silently. The pivot was correct but never announced, which on a real team reads as unexplained churn in review.
Discarded the eager-loading approach at 00:22 with no comment or message explaining why.
4 overrides of AI suggestions across the session; an independent re-review found all 4 correct.
11 prompt rewrites, each adding a constraint the AI had dropped. The usual failure mode — pasting the error back verbatim — never occurred.
Located the N+1 from the query log rather than guessing at the model layer.
Compared against 214 similar candidates
Standout Areas
Development Areas
Ask her to write the regression test for the race condition she found in this session. Tests are the one gap; watch her close it live.
Give her a PR containing a plausible but wrong AI-generated migration and ask her to review it. She scored 93 on critical evaluation — verify that transfers to reviewing someone else's AI output, not just her own.
Ask why she abandoned the eager-loading approach. The decision was right; the question is whether she can articulate it after the fact.