Ondřej Kobza
PhD Researcher & ML Engineer
PhD candidate at Czech Technical University in Prague. LLM alignment, secure code generation, reinforcement learning. 2nd place at the Amazon Nova AI Challenge.
PhD Researcher & ML Engineer
PhD candidate at Czech Technical University in Prague. LLM alignment, secure code generation, reinforcement learning. 2nd place at the Amazon Nova AI Challenge.
Focused on LLM safety, secure code generation, and alignment techniques.
My research focuses on training and aligning large language models, with the current work centered on secure and trustworthy code generation. The toolbox includes supervised fine-tuning (SFT), direct preference optimization (DPO), reinforcement learning from human feedback (RLHF), and constitutional AI methods, implemented in Python with PyTorch, Hugging Face Transformers, TRL, Verl, and Unsloth. Training runs regularly on multi-node multi-GPU clusters — Karolina (Ostrava, Czechia) and LUMI (Finland, the supercomputer of the North) — across both NVIDIA and AMD GPUs, managed via PBS and Slurm. AWS SageMaker handles additional training and deployment workloads. A core part of the work involves designing synthetic data pipelines that produce targeted training sets for LLM alignment — covering secure coding, vulnerability detection, and adversarial robustness. Before pivoting to code security, the earlier years were spent on open-domain conversational AI, building neural response generators and hybrid dialogue systems as part of the award-winning Alquist team.
O. Kobza, A. Černý, I. Dostál, J. Šedivý, M. Rigaki, M. Sladić, S. Garcia
We introduce AlquistCoder, a 3.8B-parameter coding assistant designed to generate secure code and resist adversarial manipulation. Using our constitution-guided Design–Amplify–Refine framework, we produce synthetic training data (139k SFT samples, 6.7k DPO pairs) that target vulnerability reduction and malicious-request refusal simultaneously. The model reduces statically detected vulnerability patterns and malicious-assistance rates relative to larger baselines while retaining competitive coding performance. We also release two new benchmarks — VulnBench and MalBench — for evaluating hard secure-coding and multi-turn adversarial scenarios.
O. Kobza et al. — Amazon Nova AI Challenge 2025 Proceedings
Our competition paper describing the AlquistCoder system that achieved 2nd place in the Amazon Nova AI Challenge. Presents the constitution-guided methodology for hardening AI coding assistants against adversarial red-team attacks.
J. Konrád, J. Pichl, P. Marek, P. Lorenc, O. Kobza et al. — Future Internet, Vol. 16, No. 9, 2024
Introduces the Alquist 5.0 SocialBot designed for the Alexa Prize Grand Challenge 5. Presents the integration of the novel neural response generator Barista within a hybrid architecture combining predefined dialogues with advanced neural response generation, safety mechanisms, and multimodal capabilities.
O. Kobza et al. — Future Internet, Vol. 15, No. 12, 2023
Introduces the Barista neural response generator — an enhanced architecture based on BlenderBot 3 that outperforms the original by up to 22% on specific tasks while being up to 192x faster. Presents improvements in model governance, search classification, and entity extraction for open-domain dialogue.
J. Konrád, J. Pichl, P. Marek, P. Lorenc, V. D. Ta, O. Kobza, L. Hýlová, J. Šedivý — arXiv, 2021
Describes the Alquist 4.0 system built for the Alexa Prize SocialBot Grand Challenge 4, which won 1st place. Focuses on social intelligence through generative models and dialogue personalization techniques.
Amazon organizes annual university-level AI competitions that bring together top research teams from around the world. The Alexa Prize SocialBot Grand Challenge focuses on building open-domain conversational agents, while the Nova AI Challenge targets trustworthy and secure code generation. Our CTU team has consistently placed among the top finalists.
Team Lead — 2nd Place
Led AlquistCoder to 2nd place in the international competition on building trusted, secure AI coding assistants.
Team Lead — 3rd Place
Led the Alquist team (2022–2023) in the international open-domain conversational AI competition.
Team page →
Team Member — 1st Place
Member of the winning Alquist team in the 4th edition of the Amazon Alexa Prize SocialBot Grand Challenge.
Publications in MDPI journals, co-authoring at NAACL, and CTU FEE Poster proceedings. Research spans language modeling, knowledge-preserving fine-tuning, and conversational AI. Full list on Google Scholar →
Guest appearance discussing AI research and the Amazon Nova AI Challenge results.
Interview about the AlquistCoder team's success at the Amazon Nova AI Challenge.
From industry internships to leading AI research teams.
Amazon Nova AI Challenge
Built a secure coding assistant that won 2nd place. Developed the Design–Amplify–Refine synthetic data framework and trained models with SFT + DPO alignment.
PromethistAI, Prague (part-time)
Natural language processing development.
Porsche Engineering Prague
Developed an internal AI automatization tool for data science.
Amazon Alexa Prize SocialBot Grand Challenge 5
Led the team to a top-3 finish in the international open-domain conversational AI competition.
Amazon Alexa Prize SocialBot Grand Challenge 4
Member of the winning team in the 4th edition of the competition for open-domain conversational AI systems.
CTU CIIRC E-club, Prague
Doctoral research in language modeling and LLM security. Focus on alignment techniques, secure code generation, and synthetic data methods. Thesis submission expected December 2026, graduation spring 2027.
NTT DATA
AI/ML development, cloud technologies, and IT systems integration within SAP/AWS ecosystems.
SAP Concur
Customer performance engineering.
Core technologies and areas of expertise.
PhD in Open Informatics — Language ModelingCzech Technical University, Faculty of Electrical Engineering, Prague.
Expected thesis submission Dec 2026, graduation spring 2027.
MSc in Open Informatics — Artificial IntelligenceCzech Technical University, Faculty of Electrical Engineering, Prague
Erasmus+ ExchangeKU Leuven, Faculty of Engineering, Belgium
BSc in Open Informatics — Computer SystemsCzech Technical University, Faculty of Electrical Engineering, Prague
Gymnázium UničovHigh school, Uničov, Czech Republic
Erasmus+ ProjectsParticipated in 5 Erasmus+ projects (training courses and youth exchanges) across Europe. A formative experience for building cross-cultural communication, teamwork, and social skills.
When I'm not training models, I'm training myself.
I'm originally from Uničov, a small town in Moravia, Czech Republic. I've been living in Prague since 2015 when I started my studies at CTU. Outside of research, I spend most of my free time in the mountains — alpine climbing, mountaineering, and trail running are what keep me grounded (paradoxically, often at altitude). I did athletics competitively between 2000 and 2015, specializing in the 800m run (personal record: 1:58.71), and nowadays I still enjoy running. I enjoy traveling, reading about history, and following developments in politics and economics.
Building things beyond academia.
A mobile arcade game for Android, built with React Native & Expo. Navigate through obstacles by flipping gravity — simple to pick up, tricky to master.
A Czech soundboard app for Android — featuring iconic quotes from the cult classic. Built with React Native & Expo.
Solving the E2E NLG Challenge using two transformer generative models: T5-base and GPT-2. A natural language generation pipeline for data-to-text tasks.
Whether you're interested in collaboration, have questions about my research, or just want to connect.
