Hi, I'm

Ayhan Meherrem

Computer Engineer & AI Researcher

Final year Computer Science student at Politecnico di Torino with GPA 29.42/30 or 3.93/4.00 , obsessed with generative AI from LLMs and RAG systems to the diffusion models behind image generation. I build things that push these models further, and dig into how they actually work under the hood. Currently working on implementing CariScore to get better results in caricature generation with Stable Diffusion.

View My Work Get In Touch

About Me

I'm a final year Computer Science student at Politecnico di Torino( weighted grade average 29.42/30, top 0.5% of my class) with a deep curiosity for generative AI in all its forms, from large language models and retrieval augmented systems to the diffusion models behind image generation. I'm less interested in treating these models as black boxes than in understanding how they actually work, where they break, and how to make them more reliable and controllable.

That curiosity plays out in a few directions at once: I research how to detect and attribute AI generated code, build multi agent RAG systems that reason over real world data, and explore how diffusion models can be steered including an ongoing project fine tuning Stable Diffusion to generate caricatures that stay identity consistent while pushing artistic exaggeration.

Beyond research, I love building production quality software taking an idea all the way through data, modeling, pipelines and deployment end to end. My analytical instincts go back further than my CS degree: I represented my school internationally as part of a math olympiad team, and I still approach every problem the same way break it down until it's obvious.

Experience

09/2025 — Present

AI Researcher — MALTO AI Research Laboratory, Turin

Specializing in LLM forensics and neural code authorship attribution, developing multi modal Transformer architectures to detect machine generated code and developed by which model.
Led the research team to a 10th place global rank in SemEval 2026 Task 13, engineering a robust ensemble pipeline that identifies AI artifacts across diverse programming languages.
Currently lead author of a research paper on stylometric invariance targeted at the ACL conference, aiming to establish new benchmarks in software integrity and NLP security.

06/2025 — 09/2025

Applied AI & Software Engineer Intern — Adres Gezgini Inc, İzmir

Built an NLP pipeline that assesses the semantic alignment between user queries and website content, generating AI driven suggestions when relevance was low.
Vectorized queries and sentences with Sentence Transformers (all-MiniLM-L6-v2, via Hugging Face) and spaCy embeddings, then measured semantic similarity with cosine similarity.
Developed an LLM based recommendation module using Ollama, combining prompt engineering with transformer models to propose missing topics and improvements.

Research & Publications

MALTO at SemEval-2026 Task 13: Detecting Human, AI, and Hybrid Code via Hard Negative Mining and Curriculum-Driven Ensembles

20th International Workshop on Semantic Evaluation (SemEval-2026), co-located with ACL 2026 — San Diego, California, USA, July 2026

Built a unified detection framework combining UniXcoder and CodeT5 to tell human written code apart from AI generated and hybrid code. To make the models robust against paraphrasing and adversarial rewrites, I designed a Tree sitter based canonicalization step that strips superficial variation while preserving structure, paired it with targeted data augmentation, and trained the ensemble through a 3 phase curriculum with hard negative mining pushing the models to focus on genuinely ambiguous cases instead of easy ones. The approach ranked our team 10th globally out of all submissions.

Read the Paper →

Evaluating Geometric Exaggeration and Identity Preservation in Diffusion-Based Caricature Generation

Ongoing Undergraduate Research Project, Yasar University — Advisor: Prof. Arman Savran

Caricature generation has to walk a fine line: exaggerate a face enough to read as a caricature, without losing what makes it recognizable as that person. I'm co developing CariScore, a geometric metric that quantifies this trade off directly, and using it to guide fine tuning of Stable Diffusion's cross attention layers so the model learns to push exaggeration further while staying identity consistent.

Projects

Multi-Agent Financial RAG System

An end to end system for comparative financial analysis that splits reasoning across a router agent and a synthesizer agent built on LlamaIndex, orchestrating Groq hosted Llama 3 (8B/70B) models. I parsed raw SEC HTML filings with regex to extract and tag structured content, then indexed it into Qdrant for metadata filtered retrieval. The backend runs on a secure FastAPI service with internal auth, rate limiting, and prompt injection/XSS defenses behind a Streamlit frontend, deployed via Docker Compose to Azure Container Apps with a scale to zero architecture and Arize Phoenix wired in for LLM observability and cost tracking.

LlamaIndex
Qdrant
FastAPI
Docker
Azure

Ensemble Model for Human vs. LLM Code Attribution

Given an arbitrary code snippet, this model predicts not just whether it was written by a human or an LLM, but which LLM if so. I fine tuned multiple Transformer backbones (UniXCoder, CodeT5) with hard negative mining to push the ensemble toward genuine stylistic fingerprints rather than superficial cues, on top of a tree sitter based preprocessing pipeline that canonicalizes code by generalizing function/variable names and exposing structural information.

Transformers
Tree-sitter
Hard Negative Mining
PyTorch

Semantic Relevance Optimization for Landing Pages

An end to end pipeline that measures semantic alignment between user queries and landing page content using dense vector representations instead of sparse keyword features. Built a crawler and parsing layer with Selenium/Playwright, encoded content with Sentence Transformers, and calibrated decision thresholds on real click log data.

NLP
Sentence Transformers
Vector Search
Python

Skills

Programming

AI / ML

Tools

Languages

Awards & Honors

SemEval 2026 Task 13 — Global Rank #10

Led the MALTO research team to a 10th place worldwide ranking in machine generated code detection with our ensemble pipeline.

Politecnico di Torino — Top 0.5% of Class

Achieved a 29.42/30 weighted grade average (GPA 3.93/4.00), ranking among the top performing students in Computer Science.

110L/110 con Lode — Final Examination

Achieved the highest possible grade with honors on the entrance evaluation exam.

Full Merit Scholarship — Politecnico di Torino

Awarded a full merit based scholarship (€24,000 + tuition fee reduction) for academic excellence.

Ker(PoliTo) Math Team — IMC Representative

Member of Ker(PoliTo), representing Politecnico di Torino at the International Mathematics Competition (IMC).

TÜBİTAK Mathematics Olympiad — 9th Place, Turkey

Ranked 9th nationally in the TÜBİTAK Mathematics Olympiad, Turkey's national science and math olympiad program.

Kangaroo Mathematics Olympiad — 2nd Place, Turkey

Ranked 2nd nationally in the Turkish leg of the Kangaroo Mathematics Olympiad, an international math competition held across 100+ countries.

Mediterranean Mathematics Olympiad — Silver Medal

Awarded a silver medal at the Mediterranean Mathematics Olympiad, an international competition for students from Mediterranean countries.

Turkey Mathematics Competition — Gold Medal

Awarded a gold medal at the Turkey Mathematics Competition.

Izmir Science Olympiad — Silver Medal

Awarded a silver medal at the Izmir Science Olympiad.

Get In Touch

Whether it's research collaborations, internship/job opportunities, or just to say hi — feel free to email me or reach out through the links below.

ayhan.meherrem2@gmail.com