Workopia Hire

Our vision
Our JD drafting in ChatGPT
Our Social Poster Function in ChatGPT
Tips for job posting to have best ROI based on live data

What it is

Workopia Hire is a conversational ATS — the entire hiring workflow runs inside one ChatGPT conversation. A small business owner describes a role and, without leaving the chat, gets a job description, a live apply page, a promotion kit, ranked applicants, and a signed-ready offer.

Inspiration

Hiring one person means juggling five disconnected tools: a job board, a design app, an applicant tracker, email, and a document editor. For a café or a small shop that has no HR team, that overhead is the real barrier. But the decision already happens in a chat window — so we put the whole process there.

What it does

A complete hiring loop, in five moves:

Describe the role → a complete, structured JD in seconds, grounot a blank template.
Publish (explicit approval required) → a real public apply page, instantly.
Promote → a branded QR hiring poster, a ready caption, and one-click share to 8 platforms.
Review → every applicant is auto-scored 0–100 against the role's must-haves and ranked, with the matched/missing-skill evidence and the resume attached.
Hire → interview invites, and a compliant offer letter as a real PDF — preview before send.

Built-in guardrails: market pay benchmarks (44+ countries) and st(including US state-level minimums) are shown separately, so the owner sees what the market pays and what the law requires. Below-floor offers are blocked server-side before they can be sent.

How we built it

Built with Codex + GPT-5.6 — the model plans and converses, the backend is deterministic.

Next.js 15 App Router on Vercel, live at hire.workopia.ai
An MCP server (JSON-RPC 2.0) exposing 6 tools / 19 actions, with ChatGPT Apps SDK widgets (skybridge resources) for the JD, poster, candidate pipeline, and offer preview
A self-contained OAuth 2.1 authorization server — PKCE, dynamic client registration, refresh-token rotation — so the app connects from ChatGPT (and Claude) with no external dependency
MongoDB for jobs, applications, and GridFS resumes; satori + resvg for poster rendering; Resend for email; Gotenberg for offer PDFs

Challenges we ran into

Widget mounting: ChatGPT's Apps SDK only mounts a widget when the resource is served as text/html+skybridge. A more "standards-compliant" mime silently broke every widget after the first — the fix was one line, the diagnosis was not.
Cross-tool session memory: shared-secret callers had no stable identity, so a later tool lost the applicant selected in an earlier turn. We attached a stable subject so context persists across the whole conversation.
The wage-floor guardrail: making "block below-floor offers" hold across pay units (hourly vs annual) and jurisdictions, without false positives.

Accomplishments

The loop actually closes: publish, apply, auto-score, and generate an one domain. In the demo, the guardrail blocks a real below-minimumoffer before it can be sent.

What we learned

Conversational software still needs deterministic edges — explicit approvals, server-side blocks, and evidence-backed ranking are what make it trustworthy, not just fluent.