Research

Research

Six areas of work. The first three turn on a single question: under what conditions can a generative system be relied upon?

Retrieval-Augmented Generation

Comparative analysis of RAG variants, chunking and fusion strategies, and the trade-off between retrieval and fine-tuning, evaluated on open-domain and domain-specific question answering.

RAGBM25BERT-E5Reranking

Trustworthy & Secure LLMs

Dual-LLM architectures that isolate untrusted input from privileged action, defence against prompt injection and jailbreaks, and reliability auditing of generative systems in business contexts.

Prompt InjectionLLM SecurityReliability

Domain-Specific LLMs

Assessing whether domain specialisation justifies its cost, through benchmarks of medical-domain against general-purpose models and hierarchical, intent-driven multi-agent frameworks for clinical question answering.

Medical NLPMulti-AgentBenchmarking

AI in Education

Personalised tutoring with LLMs, deployment of AI assistants inside real university courses, and the opportunities and risks generative AI brings to higher education.

AIEDTutoring SystemsChatbots

Computational Economics & NLP

Stylometric fingerprints in central bank communication, quantifying the market impact of policy statements, and threshold effects between money growth and inflation.

StylometryCentral BankingText Mining

Vision & Speech

Deep learning for flying-object classification, and cross-model knowledge distillation that improves small speech models using teacher-generated pseudo labels.

Computer VisionASRDistillation

Retrieval, reranking and generation pipeline A query is retrieved over by a sparse and a dense retriever, the candidate set is reranked, and only the reranked top-k passages reach the generator. The reranking stage is highlighted as the largest contributor to answer quality. Query BM25 BERT-E5 retrieval Reranker decisive stage LLM generation Answer top-k
Across open-domain and medical question answering, the reranking stage accounted for more of the gain in answer quality than additional pipeline complexity did. Schematic; see Şakar & Emekci (2025) and Emekci & Roxas (2026).

Projects

Four TÜBİTAK-supported research projects, two of them currently running, alongside deployed systems and a continuing research programme.

Funded Research

AI-Assisted Tax Ruling Response System (e-Özelge)

TÜBİTAK 1007

A public-sector artificial intelligence system for answering tax rulings (özelge), developed under the TÜBİTAK 1007 programme for research and development in public institutions.

Researcher · from January 2026, in progress · Principal investigator: Kasım Murat Karakaya

DataMentor

TÜBİTAK

An LLM-based platform for intelligent code generation and data analytics driven by AI agents. DataMentor extends the tutoring work of Robomentor from explanation towards execution: agents that write, run and analyse code alongside the user.

Principal investigator · from July 2025, in progress

Robomentor

TÜBİTAK 1512 · BiGG

A generative AI tutoring system developed to support interactive learning in university courses. Robomentor was integrated into Machine Learning and Data Analytics courses at TED University, providing students with continuous, personalised assistance alongside AI-augmented course material. Developed by NavigaAI Inc.

Principal investigator · January–December 2024 · Coverage: DHA · Digital Age · Gerçek Gündem

IMF Decides, Machine Learns: An AI Approach to IOs

TÜBİTAK 2232

A machine learning approach to the political economy of international organisations. The project builds a dataset of IMF programmes and models what determines loan size, the number of conditions attached, and conditionality waivers granted during implementation. A natural language processing component analyses IMF Executive Board minutes to identify alignments between country representatives and the position of the G5.

Researcher · October 2019 – February 2023 · Principal investigator: Merih Angin, Koç University

Deployed Systems

ChatTEDU

TED University

An AI-based advisory platform built to give prospective and current students immediate, accurate answers during the university admissions period. It served more than 5,000 users and, in doing so, became a testbed for LLM security: analysis of prompt-injection and jailbreak attempts observed in deployment informed a dual-LLM defence architecture in which the model exposed to untrusted input is separated from the model permitted to act.

TED University · July–September 2024 · Published in IEEE Access, 2025 · Basis of TED University’s 23rd-place WURI 2025 ranking

Dual-LLM architecture with a trust boundary Untrusted user input reaches only a quarantined model, whose output is constrained to a structured form before it crosses a trust boundary to a privileged model that holds tool and data access. untrusted zone privileged zone trust boundary User input Quarantined LLM no tool access Privileged LLM tool access Tools Data structured output only
Prompt injection reaches only the quarantined model; what crosses the boundary is constrained output rather than instructions. Schematic of the defence reported in Emekci & Budakoğlu (2025).

Research Programmes

Central Bank Communication Analytics

Ongoing

A continuing programme of work quantifying the effect of monetary policy language on financial markets, extending from the computational analysis of CBRT policy statements in my doctoral thesis to the stylometric identification of authorial signatures in policy communication and threshold models linking money growth to inflation.

Code & Open Source

Research code, teaching material and reproducibility packages on github.com/hemekci.

  • code-constrained-plan-gen Python

    Regulation-aware floor plan diffusion: a code-agnostic framework with universal guidance, demonstrated across nine building-code jurisdictions.

  • UHI Python

    Continuous morphology archetypes for urban heat mitigation across cities.

  • threshold-on-money-growth R

    Replication code for Başçı, Emekci & Apaydın (2025), Applied Economics Letters.

  • ADS525 Jupyter Notebook

    Course notebooks for Generative AI Engineering with LLMs.

  • RecommendationSystems Jupyter Notebook

    Recommendation algorithms applied to large-scale datasets.

  • PythonforFinance Jupyter Notebook

    Python for financial applications — teaching material.