Research · papers, data and open questions

Research, from production.

I work where applied AI meets labour-market data: agents that map job titles and skills to occupational taxonomies, skill extraction in lower-resource languages, and what agentic systems cost and how to evaluate them. I am open to working with research labs on these questions.

Research interests

Agents for labour-market data

Multi-agent systems that normalise free-form job titles and skills to O*NET and ESCO, with confidence gating and a human in the loop.

Skills in more languages

LLM-based skill extraction for lower-resource languages, starting with Turkish: benchmarks, annotation and where models fail.

Context engineering and cost

Token budgeting and context composition for AI coding assistants and agent harnesses: what goes in the window, and what it costs.

Evaluating agents

Evals, tracing and failure analysis for agentic systems in production, and how offline scores relate to what users see.

Papers

  1. ACM CAIS ’26 · peer-reviewedMulti-Agent Position Classification with Tool Orchestration: Use Case System for Occupational Taxonomy MappingV. Farajijobehdar, İ. Köseoğlu Sarı, N. K. Üre, E. Zeydan · Demo Track, 2026 · DOI
  2. arXiv · April 2026Tokalator: A Context Engineering Toolkit for AI Coding AssistantsV. Farajijobehdar, İ. Köseoğlu Sarı, N. K. Üre, E. Zeydan · arXiv:2604.08290
  3. arXiv · January 2026Leveraging LLMs for Turkish Skill ExtractionE. Arslan İltüzer, Ö. A. Özlü, V. Farajijobehdar, G. Eryığit · arXiv:2601.22885

Earlier: M.A. in Economics, thesis on real-time labour-market analysis; teaching assistant in Econometrics I; one patent pending.

What I bring to a lab

  • A working system. The multi-agent occupation classifier from the CAIS paper: job titles in five languages to O*NET and ESCO, with its tools and evaluation.
  • Open data. The applied AI job-ads tracker: ads read daily from 200+ public career pages and boards, classified by role, skill, task and tool, with its method, its error rate and an open dataset.
  • Open source. Tokalator, a context-engineering toolkit with 500+ installs, as a testbed for studies of token cost.
  • Industry problems. Eight years in industry and four in research: questions that come from real hiring and workforce data, and the habit of shipping evals with every system.
  • Teaching. An 18-hour applied AI course on agents, MCP and AI-native development, and university lectures on prompt engineering and tool design.

Working with a lab

I would like to spend time with a group working on labour-market and skills data, NLP for more languages, or how agentic systems are built and evaluated, for example a data science institute that works closely with industry. The length, the place (on site or hybrid) and the question are open: I would rather shape them with the host than bring a fixed plan. A good visit, for me, ends in a shared dataset, an evaluation or a paper.

Contact