AI Automation & Job Classification

The Problem

Job boards ingest data from dozens of sources — direct employer posts, ATS feeds, scraped listings, programmatic backfill. Every source uses different formatting, different title conventions, different location strings, different category labels. Without automated classification, you get duplicate listings, miscategorized jobs, broken search filters, and category pages full of irrelevant results.

Manual tagging doesn't scale. Rule-based systems break the moment a new job title variant appears. And off-the-shelf NLP models trained on generic text corpora don't understand that "Senior SWE - Platform (Remote, EMEA)" is a backend engineering role in a specific geography.

How I Work

I build custom job classification and enrichment pipelines using models trained on recruiting industry data. I've been working with AI in job tech since 2013.

Custom Job Classifiers. BERT and XLM-RoBERTa models fine-tuned for job title classification, category assignment, seniority detection, and work-type extraction (remote, hybrid, on-site). These models handle multilingual input natively — critical for European job boards operating across languages.

Job Parsing & Metadata Extraction. Structured extraction of salary ranges, required skills, experience levels, benefits, and contract types from unstructured job descriptions. This feeds your search filters, matching algorithms, and SEO-optimized category pages.

Deduplication. Fuzzy matching and embedding-based similarity to catch duplicate jobs across sources — even when titles, descriptions, and company names are formatted differently.

Vector Database Matching. Job embeddings stored in vector databases for semantic search and candidate-job matching that goes beyond keyword overlap. A "Machine Learning Engineer" matches with "AI/ML Developer" even if neither listing uses the other's exact terms.

Pipeline Integration. These components slot into your existing job ingestion flow. Ingest → clean → classify → enrich → deduplicate → serve. I handle the ML architecture; your engineering team handles the infrastructure integration.

Technologies & Tools

BERT, XLM-RoBERTa, sentence-transformers, Python (scikit-learn, PyTorch, Hugging Face), vector databases (Pinecone, Weaviate, Qdrant), custom training datasets from job board content.

Results

• Built classification pipelines processing millions of jobs daily for multiple European job boards and aggregators.
• Multilingual classifiers deployed across 10+ European languages with consistent accuracy.
• Deduplication systems reducing redundant listings by 30–40% across multi-source aggregators.

Your job data is only as good as your classification pipeline. Let's build yours.