Blazej Manczak

Building AgentWarden: (1) detect and evaluate risk vectors for AI Agents, (2) build guardrails and tooling to reduce the risk and (3) observability tool with intelligence to flag when things are going wrong. I’m hiring! Reach out if you’re excited about the safe dissemination of agentic AI systems across society, through tools that identify and manage security risks and vulnerabilities. Dec 25': Lead author of Shallow Robustness, Deep Vulnerabilities: Multi-Turn Evaluation of Medical LLMs, presented in San Diego at NeurIPS 2025

Lead and shipped 0->1 DynamoGuard LLM Output Guardrails. DynamoGuard is live in production with Experian, Lenovo, Qualcomm, and many other Fortune 500 enterprises.

Built the train & eval & inference infrastructure for training custom judge models on demand through human in the loop pipeline. Achieved pareto-frontier dominance in performance and efficiency. July '24: Lead author of PrimeGuard, presented at ICML in Vienna

- Developed (model-based) RL solutions for combinatorial problems - Applied search techniques to drive LLMs to achieve high rewards in coding problems - Published two papers on iterative improvements of language models

Worked on hierarchical reinforcement learning methods with prof. Herke van Hoof.

Developed a hierarchical reinforcement learning system capable of operating a medium-sized power grid up to 6x better than a physics-based simulator.

Building a software that automates data migration. Working on smart solutions for data mapping, data cleansing and logical reorganisation of the database schema in collaboration with a large electrical equipment producer.

- Integrated and finetuned a state-of-the-art Speech-To-Text model to the domain of online meetings. - Developed text-cleaning routines such as punctuation restoration or interjection filtering for enhanced performance on the downstream tasks. - Implemented and benchmarked extractive and abstractive summarization models against user preferences. - Deployed the models in an ML pipeline hosted on AWS.

- Decomposed video into frames and extracted the most informative features. - Performed clustering and matching of the detected features to the input query images. - Developed an open-source module for the system: https://github.com/bmanczak/AoM-LineMatching - Facilitated and consulted the app developer on the application of the system in a mobile application.

- Implemented the leading methods for face detection, CNN architectures, model validation and inference speed optimizations. - Achieved state-of-the-art results on representative facial expression recognition datasets. - Successfully deployed the model on resource-constrained drone with real-time performance of over 7 frames per second. - Presented the summary notebook at company and university ML meetups.

TA for courses: Data Analytics, Data Intensive Systems, Programming for DS, DS Ethics, Foundations of computing

Relevant courses include: Reinforcement Learning , Causality, Machine Learning I, Computer Vision I, NLP I, Deep Learning, Advanced Topics in Computational Semantics, Fairness and Explainability in AI, Deep Learning for NLP

Relevant courses include: Data Mining,Data Analytics, Statistical Computing, Data-Intensive Systems, Data Management, Business Analytics, Data Science (DS) Research Methods, Visualization, Data Structures and a series of real-life Data Challenges Math background: Calculus, Probability Theory, Linear Algebra, Numerical Linear Algebra, Numerical Analysis, Linear Optimization, Advance Mathematics for DS I and II