René Cano
All projects

No. 23rd place · LiverHack 2026

LivHire

Talent acquisition copilot · LiverHack 2026

My role
Process-domain backend · team of 4
Period
Sep 2026
Demo Code
LivHire sign-in screen: on the left the headline “El copiloto de tu proceso de talento” (the copilot for your talent process) with three benefits, and on the right the corporate account sign-in form.

The problem

At LiverHack 2026, El Puerto de Liverpool asked for a digital talent acquisition ecosystem. They recruit centrally for ~180 open roles, with an average time-to-fill of 45 days. Hiring managers can't see where their process stands or who is blocking it, candidate comparison lives in a spreadsheet, and out of ~300 candidates most never hear back.

What I built

We were a team of 4. I owned the repository and the process domain, the deterministic part of the system: 17 commits and 6,060 of 21,757 lines added (~28%, excluding lockfiles).

  • The Supabase schema: tables, 4 migrations, seed data and 44 Row Level Security policies, with an append-only audit_log.
  • The orchestrator: a 6-stage state machine from requisition to offer. The AI never rejects or makes an offer on its own: every irreversible decision is made by a person and requires a justification.
  • The business-day SLA engine: a status light per stage, a predicted fill date and escalation.
  • Role-based authentication, the app shell, the hiring manager and HRBP dashboards, and transactional email with Resend.
  • End-to-end verification scripts against the database.

My teammates built the AI layer (CV extraction, comparison with citations, a blind evaluator and a bias report), the MCP server and the Google Calendar integration.

Architecture

Architecture diagram: users (hiring manager, recruiter, HRBP and interviewers) use a Next.js app; the app calls the 6-stage orchestrator, the SLA engine and the AI layer; the orchestrator and SLA engine write to Supabase with RLS and an audit_log; the AI layer uses OpenAI; email goes out through Resend and interviews are booked on Google Calendar; an MCP server reads the same database. The parts I built are marked.

Results

We placed 3rd at LiverHack 2026. It's a hackathon prototype deployed on Vercel that you can try live. It has no usage metrics and no measured evaluation of the AI layer.

Known limits

  • The project's ~50 tests cover the AI layer and were written by a teammate. The orchestrator and the SLA engine, which are my part, have no unit tests, and there is no CI.
  • There are no usage metrics and no evaluation of the LLM extraction quality.

Links