Kaan Gümele
Applied AI Engineer & Data Scientist
Mapin Data
MSc Researcher
Istanbul University-Cerrahpaşa

Research

Topics I have worked on since starting my MSc, in approximately anti-chronological order. A complete list of publications is on ORCID.

Covert Channels in Multi-Agent LLM Systems

As LLM agents are given more freedom to reason and to talk to each other, a single compromised agent can leak task-external information to another by embedding it in stylistic, ordering or timing choices — a steganographic channel that syntactic filters and static DLP see as perfectly clean. The field has established that such covert collusion exists, but has not measured how much capacity a given interface carries. My thesis defines that quantity and draws a practical consequence from it: runtime monitoring cannot reduce the capacity; the only lever is constraining the interface at design time.

Deep Learning for Medical Imaging

On MRI, CT and X-ray, deep models can approach expert-level accuracy; the real bottleneck is not the model but labelled data. Medical annotation is expensive, slow and imbalanced — for rare classes even a few hundred examples are hard to obtain. Working with clinicians and faculty, I built transfer-learning based solutions: meningioma grading, thoracolumbar fracture detection, oral cancer classification, and an ongoing study on dental furcation. This line of work forms my academic backbone.

Autonomous Agents and Orchestration

Autonomous agents are impressive in a demo and brittle in production: they stall under ambiguity, fail silently, and are hard to debug. The question I care about is whether splitting responsibility across several agents — instead of one enormous prompt — genuinely makes the system more reliable. I built an autonomous browsing agent that substantially reduced manual QA on Google Maps, and a hierarchical orchestration system that turns natural-language input into a task plan. This line also provides the practical ground for my thesis work.

Retrieval and RAG

The hard part of RAG is not generation but retrieval: finding the right fragment in a hundred-table schema or a corpus of thousands of papers. Given the wrong context, the model confabulates confidently. I wrote a post series that walks through exactly where hallucination comes from, using a natural-language-to-SQL assistant; I am currently building an academic RAG system enriched with a citation graph to survey the literature around my thesis.


msc thesis · in progress

Measuring the Undetectable: Covert-Channel Capacity in Multi-Agent LLM Systems

Advisor: Şafak Durukan Odabaşı — Istanbul University-Cerrahpaşa

I measure the undetectable covert capacity an interface between two agents can carry in multi-agent LLM systems.

Work in progress; the paper is not published yet. Happy to share it by email on request.


publications

Titles link to the DOI record. Full list: ORCID 0009-0002-4262-0585.


education

2025 – present
Istanbul University-Cerrahpaşa
MSc, Computer Engineering
2021 – 2025
Istanbul Medeniyet University
BSc, Computer Engineering
Undergraduate thesis · June 2025 · Summary (TR)  ·  PDF