RecruitFlow

Architecture

Inspiration

Keyword filters throw out good engineers because they wrote "Postgres" instead of "PostgreSQL." Meanwhile a recruiter facing 200 resumes can't read any of them properly. Both failure modes come from the same place: nobody has encoded what actually matters for the role. RecruitFlow puts that judgement back in the hands of the person who has it: the hiring manager, and uses Qwen to apply it at scale, with its work shown.

What it does

A domain expert defines weighted criteria for a role and marks some as must-haves. For each resume, qwen-max returns a score, a rationale, and the literal evidence quote from the resume for every criterion, so every decision is auditable against the source text. Clear cases route automatically; borderline candidates (0.40–0.65) escalate to a human. A must-have scoring zero vetoes the candidate regardless of overall score, a candidate scoring 74%, above the approval threshold, is still hard-rejected if a must-have is missing. qwen-plus drafts outreach and rejection letters, held for recruiter sign-off before anything reaches a candidate.

How we built it

Node/Express agent on Alibaba Cloud ECS (Singapore), with a guarded state machine where illegal transitions throw. Cases persist in PostgreSQL 16 with full state history. Next.js dashboard served via pm2, showing the pipeline board and the per-criterion evidence table. qwen-max for screening, qwen-plus for drafting, both via DashScope. Built with Claude Code, the full build log is in BUILD_NOTES.md.

Challenges we ran into

Qwen scores relatively rather than absolutely, so calibrating the review band to hit meaningful thresholds took iteration. Rejection letters needed a compliance design: the internal scorecard never reaches the letter prompt, and only criteria scoring 6+ are passed to the model, structurally preventing the disqualifying weakness from leaking into writing, since written rejection rationales create legal exposure.

What we learned

An LLM producing a number isn't a decision system. The engineering that matters is what surrounds it, guarded transitions, deterministic business rules that can override the model, evidence grounding, and human gates at the two moments that count.