Can Ali Ateş

M.Sc. in Computer Science

Hacettepe University

🤝 Advisor: Prof. Dr. Mehmet Erkut Erdem
🤝 Co-Advisor: Assoc. Prof. Aykut Erdem
🔬 Research Interests: Multimodal LLMs

B.Eng. in Artificial Intelligence

Hacettepe University

📚 CGPA: 3.76/4.00
🥇 1st in AI Engineering
🥈 2nd in Engineering Faculty

HAVELSAN - AI Engineer

Returned to HAVELSAN as a Candidate Engineer in March and got promoted to AI Engineer in July. Contributed to the R&D project MLTrack, a no-code time-series modeling platform, by implementing state-of-the-art TSMixer and DLinear models from research papers and developing algorithms.

Currently taking a role on eye-tracking systems using Tobii Pro Glasses 3 for another R&D project in collaboration with Turkish Airlines.

Datascope - Founding AI Engineer

Joined Datascope (now Agentscope) as a founding AI Engineer. Played a key role in the early stages, got accepted into TUBITAK 1507, and became a Sabancı ARF finalist. Boosted asset growth forecasting accuracy by 60% in Sabancı DX's Industry Cycles App across multiple industries. Had to leave due to limited funding.

HAVELSAN - AI Internship (Summer)

Accepted into HAVELSAN's HIT Internship Program for the summer. Worked on predictive maintenance with time-series data and hit top 5% RMSE and accuracy on NASA's Turbofan Jet Engine dataset from Kaggle.

TUSAŞ - AI Internship (Long-Term)

Accepted into the Sky Experience Internship Program. Supported the computer vision team by labeling data and researching anomaly detection on air vehicle sensor data two days a week.

Gitek Vision - AI Internship (Summer)

Had my first startup experience in a 3-person company. Used OpenCV for image processing, labeled apples and tomatoes manually then detect them with YOLOv3, and built a vision algorithm from scratch for industrial screw detection with 90%+ accuracy.

TUSAŞ - AI Internship (Long-Term)

Earned an early long-term internship opportunity by ranking in the top 5 GPA of my department in the second year. Researched on reinforcement learning basics one day per week.

RAD-ACE: Multimodal Large Language Models as Radiology Assistants

A continuation of the CE-MedAI project. Built RAD-ACE-CoT, a 16K-pair dataset of radiology images with CoT diagnostic reasoning. Fine-tuned Qwen2.5-VL (3B & 7B) and LLaMA 3.2-Vision 11B using Unsloth with LoRA adapters for efficient multimodal training. Developed a radiology-specific evaluation pipeline leveraging the LLM-as-a-Judge methodology across multiple clinical reasoning metrics.

CE-MedAI: Domain-Specific Multimodal LLMs for Radiology

My first step into the world of Multimodal LLMs. Fine-tuned CLIP and RAD-DINO as visual encoders, and OPT125M, Qwen2.5 0.5B Instruct, and TinyLLaVa 1.5B as language models. The end result was a radiology-focused MLLM agent that can generate structured medical reports from medical images using alignment and instruction tuning.

Smart Fridge

My EU-funded graduation project, done with Hacettepe University's Food Engineering Department and advised by Prof. Dr. Mehmet Erkut Erdem. The goal was simple, cut down food waste by tracking how fresh fruits and veggies are. Fine-tuned YOLOv8-S to detect them and used ResNet-101 to estimate freshness and analyze decay.

Industry Cycles App

A Sabancı ARF side project where I worked with private Sabancı DX data covering quarterly industry cycles in the cement, electricity, technology, tire, and insurance sectors. Improved one- and two-quarter ahead forecasting accuracy by 60% using machine learning, deep learning, and time-series models.

Neuro Deep Advisor

This project holds a special place for me — one of our team members was diagnosed with cancer during the process, and thankfully, he's healthy now.

Developed a decision support system to classify Alzheimer's disease severity from brain MRI scans. Built and ensembled six custom CNN architectures along with YOLOv5 and YOLOv8 models, then integrated the system into a GUI for real-time prediction and visualization.

Project LEAFS

My first end-to-end project, focusing on tracking and assessing student attitudes during lectures. Created a comprehensive dataset by capturing class images by hand and Google Images scraping. YOLOv5 is fine-tuned with this manually labeled dataset and integrated into a GUI for real-time class monitoring.

Surfing the Bitcoin Waves

Collected Bitcoin's time-series stock market data from private Hyblockcapital API. Compared different models to observe effects of various trader types (whales, bots, top traders) over the Bitcoin stock market.