HR-FlowAgent

Autonomous Hiring. Zero Friction.

Comment

Inspiration

We were tired of seeing recruitment teams drowning in CVs. Seriously—weeks of screening, scheduling nightmares, back-and-forth emails. When we saw the "All Things Agentic" hackathon, we thought: what if we just... let AI agents handle it? Not as a gimmick, but as a real system that actually works.

What It Does

HR Flow Agent automates the entire recruitment pipeline:

  1. Ingest — Candidates submit CVs and applications
  2. Screen — AI filters based on actual job requirements
  3. Assess — Technical evaluations and questionnaires happen automatically
  4. Schedule — Interviews get booked in Google Calendar without the back-and-forth
  5. Onboard — New hires get everything they need without manual setup

It's 5 specialized agents working together, powered by Gemini 3.5, talking to real Google Cloud services. No mock data, no shortcuts.

How We Built It

We split the work:

Agent Core — 5 independent agents, each with one job. They use Gemini 3.5 with strict schemas so they don't hallucinate. Built-in repair loops catch mistakes and fix them.

Backend Server — Express that orchestrates everything. State machine with 14 states keeps things sane. Real integrations with Firestore, Gmail, Calendar, Forms—not mocked.

Frontend Dashboard — For hiring teams to see what's happening. Plus we added Veo to generate personalized welcome videos for candidates (bonus points).

Testing — 23 tests that actually pass. We tested the real GCP services, not just happy paths.

Challenges We Ran Into

Getting 5 agents to coordinate without stepping on each other's toes was harder than expected. State management across async operations is tricky.

OAuth flows with Google services? Nightmare. Service accounts, refresh tokens, rate limits—we had to learn it all.

The agent runtime needed repair loops because Gemini sometimes returns slightly malformed JSON. We built error recovery that actually works.

Scheduling interviews without double-booking was a puzzle. Google Calendar API has quirks.

Implementing Cloud Run with proper secrets management took longer than coding the features.

Accomplishments We're Proud Of

It actually works end-to-end. No fake data, no shortcuts.

Real GCP services—Firestore, Gmail, Calendar, Forms all integrated.

23 passing tests. We tested the hard parts.

Clean code. Ports pattern, state machine, proper error handling.

Bonus features working—Veo integration for welcome videos.

Production-ready. Docker, logging, input validation, deployment scripts.

We learned a ton about agentic systems and actually built something useful.

What We Learned

Agentic patterns actually work. Multiple agents with clear responsibilities, coordinated through a state machine—it's elegant and maintainable.

Gemini 3.5 is solid. Good reasoning, reliable JSON output, handles complex workflows without breaking.

Google Cloud services are powerful when you understand them. The integration work paid off.

Testing autonomous systems is non-negotiable. Without tests, debugging agent behavior is impossible.

State machines make complexity manageable. 14 states and 6 events sounds like a lot, but it's way cleaner than spaghetti logic.

What's Next for HR-FlowAgent

Real companies need more. Multi-language support for global hiring. Analytics dashboard showing time-to-hire, candidate quality, and pipeline health. Integration with real ATS systems (Workday, Greenhouse). Feedback loops so the system gets smarter over time.

We're thinking about Gemma fine-tuning for domain-specific screening. Mobile app for hiring managers on the go. Real-time collaboration features.

But first—let's see if this actually solves the problem.

Built With

Share this project:

Updates

Submission history