The Agent parses each resume in an isolated sandbox, scores it against an explicit rubric you co-design, and ranks them. Every score traces back to the exact bullet point in the resume — no "the model said so". The rubric is saved to your Drive as a `rubric.yaml`; the next intake reads the same file and uses the same weights.
No "the model said so". Each dimension's score quotes a literal bullet from the candidate's resume. The rubric is co-designed; you set the weights.
“Built and scaled the payments routing service at Stripe (2019–2023)”
“Lead author of the gRPC mesh RFC adopted org-wide”
“7 years on payment infrastructure at scale”
“Mentored 4 engineers · scope: 30M req/day”
“Conference speaker (3× QCon)”
“Author of an open-source Rust queue library used in production at 3 firms”
“Limited cloud-native operator experience”
“5 years on distributed systems”
“Tech lead, 8-person team, $4M ARR product”
“Strong written, modest spoken evidence”
“Comfortable across Go and Rust; broad foundation”
“Production-grade observability stack experience”
“4 years, partly in research environment”
“Mostly IC at boutique firm; smaller scope”
“Active blogger; clear technical writing”
The 5 dimensions + weights + flag rules are written to a rubric.yaml in your Drive's hiring folder. Next batch, the Agent reads the yaml first and scores against it — HR doesn't re-explain the role every Monday.
role: sr-backend-engineer weights: must_haves: 0.35 nice_to_have: 0.10 years: 0.15 scope: 0.25 communication: 0.15 flags: - claim_mismatch_team_vs_ic
JD as PDF or DOCX, resumes as a folder or zip. The Agent parses every file in the sandbox, including weird formats (image-only PDFs get OCR'd, DOCX with embedded tables gets extracted properly).
Don't trust a black box. The Agent proposes 5 dimensions and weights, you adjust in chat ("raise must-have skills to 35%, drop years to 10%"). Scoring only starts once you sign off. The rubric is written to `rubric.yaml` in Drive so the next intake for the same role reads it back automatically.
The .xlsx lists every candidate with per-dimension scores and a one-sentence citation per dimension (the literal bullet from the resume that supported the score). Top-10 detail PDF goes deeper. Flagged inconsistencies (e.g. "led 50-person team" but listed as IC) are surfaced — you decide.
The .xlsx has a "why this score" column next to every dimension. The Agent quotes the literal sentence from the resume that justifies the score. If you disagree, you click into the source. There is no "the model said so" — every decision is auditable.
You see the proposed rubric, you change the weights, you sign off — only then does scoring start. Compare with most ATS scoring tools where the weighting is opaque and the model picks for you. Here the rubric is the contract between you and the Agent.
The rubric is written to a `rubric.yaml` file in your Drive's hiring folder ("Sr Backend Engineer · 5-dim · 2026-05 weights"). Next time a candidate batch comes in, the Agent reads that YAML before scoring — same dimensions, same weights, same flag rules. The HR team doesn't re-explain the role every Monday.
Math errors: "led a team of 50" but their job title is IC for that period. Tenure puzzles: "promoted to staff in 2023" but resume says staff since 2021. Skill ghost-claims: lists Rust as a top skill but no project ever used it. The Agent surfaces these as flags — you decide if each is a false positive or a real concern.
The Agent detects the resume language and scores in that language against your JD (also language-detected). Cross-language scoring (Chinese resume vs English JD, for example) is fully supported — no need to translate first. You can ask the Agent to surface the original-language sentence + an English gloss in citations if the hiring team is multi-lingual.
Sign in, hand it over to the Agent — the finished file lands in your Drive.