Yusuf Erdem Erdem

LLM & VLM Researcher at Baykar · NLP Research Fellow at YTU COSMOS

Yusuf Erdem Erdem

I build language and vision-language models, from data and pretraining to deployment.

At Baykar, I work on domain-specialized LLMs and VLMs, from agentic tool calling to onboard aircraft. At YTU COSMOS, I develop state-of-the-art open-source Turkish LLMs, VLMs and embedding models and novel training and inference methods.

B.S. Computer Science, Yıldız Technical UniversityHigh Honors, GPA 3.72/4, July 2026

See my experiencePublications
Yusuf Erdem Erdem

Work Experience

Research roles across an academic lab and industry, from open-source Turkish models to on-device and domain-specialized systems.

Oct 2025 - Present

Baykar Technology

LLM & VLM Researcher

Domain-specialized vision-language and language models for onboard and on-device deployment; agentic tool calling; evaluation of Turkish LLMs.

  • Integrated a domain-adapted VLM directly onto UAV's, enabling decision-making at the edge.
  • Shipped an on-device mobile LLM agent with tool calling company-wide, helping task management and internal workflows.
  • Architected a fully autonomous LLM news agency, automating the process from source collection and evaluation to scheduled publishing.
  • Conducted LLM/VLM training at scale, including large-scale dataset generation and preprocessing.

Jul 2024 - Present

YTU COSMOS AI Research Lab

NLP Research Fellow

Open-source Turkish language, vision-language and embedding models; efficient inference; large-scale data curation; distributed training on HPC clusters.

  • Developed state-of-the-art open-source Turkish LLMs, VLMs and embedding models.
  • Curated and published large-scale open-source textual and multimodal datasets with multi-million examples.
  • Created ChainFlow, a novel flow matching speculative decoding method surpassing Eagle3 and DFlash in vLLM backend.
  • Research grants from the Barcelona Supercomputing Center (BSC), Türkiye's HPC Center and national funding.

All work experience →

Research

Three threads: building the models, teaching them to reason and specialize, and making them run where they're needed.

All research and publications →

Projects, Models & Datasets

Models, datasets, demos and systems I've built or co-built, mostly open source.

Selected models

ModernBERT-TR

A long-context (8,192 tokens) Turkish encoder with a family of specialized downstream models.

HF ↗Blog ↗

ChainFlow

A flow-matching drafter for lossless speculative decoding in vLLM: up to 2.5x faster on Qwen3.5 (4B to 27B), with Turkish-adapted checkpoints.

HF ↗Code ↗Project site ↗

Cosmos-LLaVA

Turkish's first large-scale multimodal datasets along with a vision-language instruction-following model.

HF ↗IEEE ↗

Selected datasets

ABSA-TR

Turkish Aspect-based sentiment analysis dataset. Built from native sources in the e-commerce, supplements, and movie domains.

HF ↗

Guardrail-TR

About 405K Turkish prompts for safe/unsafe classification. Multi-label hazard categories covering 11 labels.

HF ↗

Turkish-LLaVA-Finetune

Turkish visual instruction fine-tuning dataset. One of the datasets behind Cosmos-LLaVA.

HF ↗IEEE ↗

All projects, models and datasets →

Leadership, Community and Honors

Organizing, teaching and mentoring around Türkiye's AI community, and the honors along the way.

Leadership and community

Talks

NVIDIA developer community managers

NVIDIAJun 2026

Spoke about where AI stands in Türkiye and how COSMOS fits into it, sharing work from the lab and my own.

Talks

Hugging Face Builders

Open-source AI meet-upMay 2026

Presented COSMOS's work and mine to the open-source AI community.

Organizing & events

TEKNOFEST

T3 · Baykar TechnologySep 2023 - Oct 2026

Took part in TEKNOFEST, Türkiye's national technology festival, as a representative of the T3 Foundation and of Baykar.

Honors & awards

  • 1st Place, TEKNOFEST Turkish Natural Language Processing Competition

    The largest national NLP competition. Built a legal decision-support system: a Turkish-law-specialized embedding model, agentic legal retrieval and reasoning, and a graph-based document analysis tool.

  • Best Paper Award, UBMK 2025

    For Improving reasoning abilities of large language models via logic games.

View all →

News

Recent milestones, talks and releases.

  • Presented my B.S. thesis work, Recency-Aware Session Representations for Real-Time Purchase Propensity Prediction, at UBMK 2026. The study was supervised by Prof. Mehmet S. Aktaş.

    PaperTalkSupervisor ↗

  • Graduated from Yıldız Technical University in Computer Science with High Honors (GPA 3.72/4). Thesis on user navigation behavior modeling became UBMK 2026 paper.

    Education

    Group of five in graduation gowns, most holding diploma scrolls, in front of a university building

  • Spoke with NVIDIA's developer community managers about where AI stands in Türkiye and how COSMOS fits into it, sharing work from the lab and my own.

    Talk

    Group photo of presenters and hosts in front of a screen showing the slide "Cosmos da Kimmiş?"

  • My thesis was chosen among the top 10 projects of 2026 in the YTU Computer Science department. I presented it to the faculty and jury.

    AwardEducation

    Large group posing with certificates on stage under a "Demo Day" screen at a graduation projects exhibition

  • Presented COSMOS's work and mine at Hugging Face Builders, an open-source AI meet-up for the AI community.

    TalkCommunity

    Group photo of about a dozen people in front of a presentation slide that reads "Cosmos da Kimmiş?"

View all →

Get in touch

Questions about my work, papers or possible collaborations are welcome.

More ways to reach me →