vesna

Sample output

One backend role, start to finish.

No login, no marketing screenshots — just what Vesna produces for a single role. Lightly anonymized; the structure is exactly what a design partner sees.

A rubric is a few building blocks

A good rubric isn't a vibe — it's a few typed, weighted blocks the recruiter sets once. Vesna scores every applicant against exactly these, not a generic idea of "good."

The role
Senior Backend Engineer — Payments

Owns the backend behind payments — high-throughput APIs, settlement, reconciliation.

Must-have pass / fail

Has built and operated LLM/agent systems in production — tool-calling, evals, cost control, real failure modes.

Nice-to-have adds points
  • Payments / fintech / settlement experience
  • Owned a backend service end-to-end — on-call, runbooks, incident response

Same on paper. Different outcome.

Two candidates, both strong-looking, both used AI. Read down the columns: they start identical — then Vesna tells them apart on substance.

Candidate Aauthentic
Candidate Bai-polished
On paper
Looks strong
Looks strong
Used AI
Yes
Yes
↑  identical so far — now Vesna reads for substance  ↓
Substance
Real — and backed up on paper
Sounds real — but can't be checked on paper
Vesna's move
Advance — nothing to ask
Ask one question first
Outcome
✓ Goes to the human
→ Shown to the human, with the gap

Candidate A — the substance is real

Vesna pulls out the evidence that matters and reads it. This candidate used AI to tighten their résumé and still comes back authentic — because the work is real.

Candidate A
Senior Backend Engineer — Payments
authentic
Evidence found · mapped to the must-have
Shipped a tool-using support agent (6-tool action layer) handling ~3,100 tickets/week unattended; drove per-resolution cost from $0.34 to $0.11.
Built the eval harness — 240 graded transcripts, run on every prompt/model change; caught a tool-selection regression a vibe-check missed.
Polish
authentic — substance over polish
Must-have
backed up on paper
Outcome
✓ advances to the human's queue

Candidate B — looks the same, so we ask

Looks just as specific and confident — but the must-have can't be checked on paper. Vesna doesn't guess and doesn't reject. It asks one question, by email.

Candidate B
Senior Backend Engineer — Payments
ai-polished
The question · sent to the candidate (magic-link, no login)

QYou mention production agent work — can you describe one concrete system: what it did autonomously, a real failure mode you hit, and how you measured whether it worked?

A"In production I handled a cascading retry storm at 2am by adding a circuit breaker on the planner step. We kept things stable after that."
The read
generic retry/queue work — names no agent system, eval, or autonomous behavior
Outcome
→ shown to the human with the gap, not advanced

The recruiter makes the call — Vesna never decides. But it turned an un-answerable "is this real?" into a five-minute, defensible human judgment, with the candidate's own words on the record.

The audit trail

Every application carries a timeline that can't be quietly edited. "Why did we pass on Candidate B?" is one screen — Vesna's steps in rust, the human's own actions in black.

10:02
Application received
10:02
Evidence extracted · 6 items
10:03
Scored · polish ai-polished · must-have unsupported
10:03
Routed to clarification · must-have gap
10:05
Question approved & sent · recruiter
14:18
Candidate responded
14:18
Response evaluated · gap not closed
15:40
Decided · hold · "no concrete agent system; revisit if pool thins" · recruiter

That's the product

A polished pool, sorted on substance; the cases that can't be checked handed to a human with the candidate's own words; every decision defensible. Want to run it on your roles?

Book a 15-min conversation What the pilot involves