How we turned a staffing firm's manual workflow into a production-grade AI talent engine.
A production AI talent engine for a staffing firm: extract, score and anonymize candidates in seconds.
The project
Sobit Technologies is a Spanish staffing and recruiting firm. Their day ran on LinkedIn, hand-anonymized CVs and candidates emailed one by one, with no searchable pool. We designed and built a unified web platform — internal CRM, client portal and resource portal — with an AI layer at its core that handles the mechanical work so recruiters can focus on people.
The challenge
Every candidate went through the same manual funnel: search LinkedIn, download the CV, anonymize it by hand (10-15 minutes per document, stripping names, companies and dates) and email it to the client. Matching candidates against each opening was subjective and lived in the recruiter's head.
Moving CVs full of personal data around was also a GDPR risk. And without a central, searchable candidate pool, every new search started from scratch.
Our approach
We designed and built a multi-area platform on Supabase and React, and put a production AI layer — not a demo — at its core to automate exactly the work that was slowing the team down.
- CV extraction: 20+ structured fields (education, experience, technologies, seniority, salary expectations, visa…) from any PDF in ~1 second.
- Explainable candidate-to-opening scoring, with the reason behind each match.
- Automatic CV pseudonymization (PDF redaction) with preview and manual override.
- Semantic search across the entire candidate pool via vector embeddings.
- Intelligent deduplication of repeated candidates in the database.
- AI-generated LinkedIn Boolean queries and outreach templates from the job description.
- Client portal with anonymized profiles, interview requests and private notes.
- GDPR by default: candidate erasure workflow and EU data residency on every AI call.
The product
How we built it
Every AI call routes through Vercel AI Gateway with EU data residency and two providers: Claude Haiku 4.5 for the heavy tasks (extraction, scoring, deduplication) and Gemini 2.5 Flash for the light ones (queries and templates). Semantic search runs on pg_vector with HNSW indexes inside the same Postgres.
We treated it as production software, not an experiment: provider fallback chains, a circuit breaker for outages, prompt versioning with an audit trail, per-call cost observability and multi-role RLS. The result is a reliable AI layer that costs on the order of €20/month at the projected scale.
Results
Sobit went from a manual, scattered process to a single platform in production, shipping in phases. The mechanical work that used to eat the team's afternoons now happens in seconds and is fully traceable.
Let's talk about your software.
30 minutes, no strings. You'll leave with a clear idea of how we can help.