JobPilot

Qwen Cloud Autopilot that searches jobs in your Chrome, scores fit against CV + GitHub, and drafts a suggested CV only after human approval.

Comment

Inspiration

I’m an AI / software engineer. Job hunting while still learning and shipping felt twice as heavy. The worst part was not finding roles but rewriting the same CV for every job: swapping projects and bullets so real work finally “matched” a new description.

I wanted an Autopilot that searches sources I already use, scores fit against my CV + GitHub, proposes project swaps, and only generates a tailored CV after I approve. That is why I built JobPilot with Qwen Cloud for Track 4: Autopilot Agent.

What it does

  1. Upload CV + connect GitHub
  2. Cloud Qwen drives LinkedIn Posts search; a desktop Helper runs the clicks in your logged-in Chrome via Kimi WebBridge
  3. Cloud agents score each job and propose keep/swap plans
  4. You approve; JobPilot drafts a suggested CV .docx (same layout; does not overwrite your master CV)

Loop: search → score → approve → suggested CV. Agents propose. Humans decide.

How we built it

Three tiers: Alibaba ECS (React, FastAPI, LangGraph, all Qwen calls) · Search Helper on Windows · Chrome for personal sessions.

Qwen via DashScope compatible-mode:

  • qwen-turbo: CV skills
  • qwen3.7-plus: browser ReAct + evidence
  • qwen3.7-max: scoring, swaps, suggested CV
  • embeddings + rerank: GitHub evidence retrieval

Qwen keys stay on ECS. Helper is browser hands only.

Challenges we ran into

Separating cloud brain from local browser hands; keeping structured JSON reliable under parallel job analysis; shipping on Alibaba ECS so personal Chrome never had to live in a datacenter.

Accomplishments that we're proud of

End-to-end Autopilot on Qwen Cloud + Alibaba ECS; HITL before any suggested CV; right model per step; public MIT repo + live demo.

What we learned

Put Qwen on the backend. Separate reasoning from irreversible actions. Deploy early. Write the trust rule first, then the graph.

What's next for JobPilot

I plan for JobPilot to work across a developer’s favorite job-hunting sites, not only LinkedIn Posts. Because JobPilot is built for developers, users should be able to customize which sources and surfaces they use, while the same three-tier loop stays: Qwen reasons in the cloud, the Helper acts in their Chrome, and they still approve before any suggested CV.

Built With

Share this project:

Updates