findjobs-agent 1.1.0

---

What is FindJobs-Agent?

A full-stack job search assistant that crawls postings from major tech companies, analyzes them with LLMs, parses your resume, and runs AI mock interviews — so you can focus on preparing, not sifting through job boards.

The whole loop in one terminal run, offline: 54 sample postings crawled and structured, resume×job skill gaps, and a mock interview with a feedback report. Replay it with python examples/full_journey_demo.py.

Try It in 2 Minutes (Demo Mode)

No API key? No problem. Demo mode runs the whole app offline:

pip install -e .
FINDJOBS_DEMO=1 python -m findjobs.api_server   # backend on :5000, seeds sample jobs on first run
cd FrontEnd && npm install && npm run dev   # frontend on :5173

With FINDJOBS_DEMO=1, the server fills an empty jobs.db from data/sample_jobs.json (54 hand-written postings from the companies the crawlers target) and routes every LLM call through demo_llm.py, a deterministic stub that answers from the prompt content and never touches the network, even if a key is configured. Upload the bundled data/sample_resume.pdf on the resume page to walk the full loop: parsing, match scores, a 3-stage mock interview, and the tracking board. While the backend runs in demo mode, a small "Demo" badge shows in the navbar.

Both fixtures can be regenerated with python scripts/seed_demo_data.py and python scripts/make_sample_resume.py.

How It Works

Four pieces wired into one flow: a crawler pulls postings from company career sites, an LLM reads each one for requirements and skills, your resume gets parsed and scored against them, and any posting can drive an AI mock interview straight off its job description. The React frontend ties it together, so you go from "what's out there" to "let me practice for this one" without leaving the app.

Features

• Job crawler — pulls postings from Tencent, NetEase, ByteDance, Amazon and more, via API or Selenium, with automatic cleaning and normalization.
• LLM analysis — extracts education/major requirements, scores skill tags (1–5), and classifies each posting into a job taxonomy.
• Resume parsing & matching — parses PDF/Word resumes, scores skills, and computes a case-insensitive job-resume match percentage.
• AI mock interview — generates questions from any job description and runs a multi-turn interview with real-time feedback.
• Demo mode: set FINDJOBS_DEMO=1 to run the full app offline with seeded sample jobs, a bundled sample resume, and a deterministic stub LLM instead of a paid key.
• SQLite persistence — analyzed postings are stored in a local jobs.db; existing CSV data is migrated automatically on first run, with CSV/JSON as fallback. Uploaded resumes and mock-interview transcripts live in the same database, so restarting the API server no longer wipes them.

Project Structure

FindJobs-Agent/
├── FrontEnd/                # React frontend
│   ├── src/
│   │   ├── components/      # Page components
│   │   │   ├── JobsPage.tsx       # Job browsing
│   │   │   ├── ResumePage.tsx     # Resume analysis
│   │   │   └── InterviewPage.tsx  # AI interview
│   │   └── App.tsx
│   └── package.json
├── findjobs/                # The package: crawler, matcher, resume, interview, pipeline
│   ├── pipeline.py          # Main pipeline; also the `findjobs` CLI entry
│   ├── job_agent.py         # LLM job analysis agent
│   ├── api_server.py        # Flask API server
│   ├── storage.py           # SQLite store: jobs, applications, resumes, interviews (jobs.db)
│   ├── interview_agent.py   # AI interview module
│   ├── resume_parser.py     # Resume parser
│   └── ...                  # crawlers, scoring, LLM client, demo stub
├── data/                    # Label library + taxonomy + demo fixtures
├── config/                  # llm_config.json and deployment configs
├── docs/                    # API reference and project notes
├── scripts/                 # Helpers: demo seeding, sample resume generator, crawler smoke test
├── tests/                   # pytest suite
├── pyproject.toml
└── requirements.txt

Quick Start

Prerequisites

• Python 3.10+
• Node.js 18+
• Chrome (required for Selenium crawler)

1. Clone the repo

git clone https://github.com/he-yufeng/FindJobs-Agent.git
cd FindJobs-Agent

2. Install the backend

pip install findjobs-agent   # straight from PyPI
# or, for development from this repo:
pip install -e .

3. Set up your API key

Create an API_key.md file with your OpenAI API key:

sk-your-api-key-here

4. Start the backend

python -m findjobs.api_server

5. Start the frontend

cd FrontEnd
npm install
npm run dev

6. Open the app

Visit http://localhost:5173 in your browser.

Data Pipeline

findjobs chains crawl → analyze → score → serve. Run the whole thing, or a single stage:

findjobs                                                        # crawl + analyze + build site data
python -m findjobs.job_crawler_v2 -c tencent netease amazon -m 300   # crawl only (--list shows companies)
findjobs --analyze-only --max-jobs 50                           # analyze only (for testing)

An optional aggregator source is available via freehire.me, an open IT job board whose public API needs no key. Each result already carries the full markdown JD and the employer's own ATS apply link, so there is no detail page to re-fetch. It stays off by default; flip it on with --freehire, or run it standalone:

findjobs --freehire --freehire-query "backend"                    # company crawlers + freehire
python -m findjobs.freehire_source -q "ml engineer" --skills python,pytorch --countries us,de -m 100
python -m findjobs.freehire_source --list-facets skills           # live filter vocabulary (skill slugs, country codes)

Pluggable job sources. Crawlers, third-party APIs, and local JSON files all enter through one JobSource interface (findjobs/job_source.py): implement name + fetch() returning canonical job dicts, then register_source(...) — no pipeline edits. The pipeline walks the registry (company crawlers by default, freehire when enabled), dedupes across sources, and pipeline.step1_crawl_jobs(sources=[...]) accepts a full replacement list for custom feeds.

A third bundled source, Remotive, widens coverage to general remote jobs (their public API also needs no key). Unlike the aggregator adapters it is a direct JobSource with no subprocess: instantiate it yourself or swap it into the registry:

from findjobs.job_source import register_source
from findjobs.remotive_source import RemotiveSource

register_source(RemotiveSource(search="llm", category="software-dev", limit=100))
# or take the whole pipeline over with a custom list:
# pipeline.step1_crawl_jobs(sources=[RemotiveSource(limit=100)])

Filter vocabularies are read from /api/v1/jobs/facets at runtime instead of being hardcoded.

Job data sources

The crawler registry only lists adapters verified live (--list), all fetched through the public job APIs that each careers site serves to job seekers. No login, no key, no page scraping behind a wall.

Last full verification 2026-10-13 (python scripts/verify_crawlers_v2.py reruns it on demand): six targeted crawlers are live, plus the Remotive remote-jobs source (verified 2026-09-17):

Source	Endpoint	Notes
Tencent (腾讯)	careers.tencent.com query API	up to 400 per pull
ByteDance (字节跳动)	jobs.bytedance.com portal API	moved to the current POST payload; the old GET form is gone
NetEase (网易)	public search API	social + campus
Amazon	amazon.jobs search API	global, CN filter supported
DJI (大疆)	Moka ATS (apply.careers.dji.com)	social; responses are AES-encrypted, decrypted locally (pycryptodome)
Ctrip (携程)	Moka ATS (app.mokahr.com)	campus channel; adapter verified healthy, zero openings as of 2026-10
Remotive	remotive.com/api/remote-jobs	key-less remote-jobs API; full JD (HTML stripped to text) and apply links included

data/latest_jobs.json holds the latest snapshot: 1,180 postings across these four (refreshed 2026-09-24), committed so the pipeline has real data to chew on without a fresh crawl. data/latest_enriched.jsonl is the same set after LLM analysis (degree, major requirement, scored skill tags, job family), so the site and the matcher work out of the box with zero API spend.

Skill tags only mean "the model found a genuine match in the technical label library": planning/design/operations jobs are never force-labeled, so an empty tag list is a legitimate result that reads as "non-technical role", not missing data.

Alibaba, Meituan and JD stay in the registry as Selenium adapters and want a local Chrome plus webdriver-manager. The other 2023-era adapters (Baidu, Kuaishou, Xiaomi, Bilibili, DiDi, Pinduoduo, Huawei, NIO, XPeng, Li Auto, OPPO, VIVO, SenseTime, MiHoYo, SHEIN, Shopee, KE, Yuanfudao, Zuoyebang, Zhilian, Lagou, Microsoft, Google) all failed the 2026-10-13 verification, mostly WAF gates or moved endpoints, and were removed from the registry; their code stays in git history as repair reference, and an issue asking for one moves it up the queue.

API Endpoints

Endpoint	Method	Description
/api/health	GET	Health check
/api/resume/upload	POST	Upload a PDF resume, returns parsed info + skill scores
/api/resume/<id>	GET	Resume details
/api/resume/file/<id>	GET	Original PDF
/api/jobs	GET	List job postings
/api/jobs/match	POST	Match a resume against jobs ({"resume_id": "..."})
/api/interview/start	POST	Start a mock interview, returns session_id
/api/interview/<id>/message	POST	Send an answer, get score + next question
/api/interview	GET	List interview sessions
/api/interview/<id>	GET	Session detail with messages
/api/applications	GET	Application board: status per tracked job
/api/applications/<job_id>	PUT	Set or update an application's status
/api/applications/<job_id>	DELETE	Remove a job from the board

Roadmap

Crawl, analyze, resume match, mock interview, and the application board all work end to end. The next steps widen the funnel and get closer to a real screen:

• More job sources — extend the crawler beyond the current company set to job boards and aggregators, so matching isn't limited to a fixed list.
• Incremental crawls — track which postings were already seen and fetch only new ones, instead of re-crawling and re-analyzing the full set each run.
• Voice mock interviews — speech in and out for the AI interviewer, closer to a real screen than a text chat.

Related Projects

FindJobs-Agent is one of the applied agents I've built. A few others you might find useful:

• CoreCoder — want to understand how a coding agent really works? Read the whole ~1k-line engine end to end, not a black box.
• RepoWiki — dropped into an unfamiliar codebase? It gives you a guided wiki and a where-to-start reading path, a self-hostable DeepWiki alternative.
• ContractGuard — catch the risky clauses before you sign: it reads contracts and flags the dangerous bits.
• GitSense — want to contribute to open source? It finds issues worth your time and gauges whether your PR will get merged.
• CodeABC — understand any codebase even if you don't code, built for non-programmers.

Contributing

Issues and pull requests are welcome!

License

MIT License