Skip to content
View bmanczak's full-sized avatar
🎯
Focusing
🎯
Focusing

Block or report bmanczak

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Please don't include any personal information such as legal names or email addresses. Maximum 100 characters, markdown supported. This note will be visible to only you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
bmanczak/README.md

Hi there 👋!

I'm a lead research engineer at Dynamo AI (YC 22) based in Amsterdam (NL) / Gdynia (PL). I work on (synthetic) data flywheels, evaluations, and training, all aimed at building efficient and aligned custom guardrailing and judge models. The complexity—and fun—lies in tackling subjective, under-specified objectives through iterative human-model alignment.

Before joining Dynamo I worked in RL for Combinatorial Optimization and Code Generation teams at Qualcomm AI Research in Amsterdam. I studied Artifical Intelligence at the Univeristy of Amsterdam, specializing in Reinforcement Learning where I did a 9 month intership at Amsterdam Machine Learning lab with prof. Herke van Hoof.

Projects I'm particularly proud of:

  • 🌟 Built Dynamo's output guardrail offering and team from scratch into a mature, high-demand product. I touched every part of the stack—from defining evaluation sets with PMs, setting up annotation and feedback loops, synthetic data generation, to training and implementing post-training interventions for efficient inference. The product now safeguards AI deployments at several Fortune 500 companies (1, 2, 3).
  • 🚀 With my team at Qualcomm, we achieved SOTA on The Abstraction and Reasoning Challenge (ARC) using a ~220M parameter language model by combining hindsight relabeling and prioritized hindsight replay (ICML '24 paper). Also proud of exploring MCTS as a neurally-guided decoding strategy for zero-human-data regimes, despite it turning into a valuable learning experience (ICML '24 workshop paper).
  • ⚡ Demonstrated that hierarchical RL can alleviate congestion in power grids up to 6x more effectively than physics-based simulators, confirming the advantage of hierarchical policies. We shared these insights in a paper.

Outside work, I’m passionate about endurance sports 🏊🚴‍♂️🏃‍♂️ and the science behind peak human performance. My favorite is Middle Distance Triathlon (70.3 Ironman), and I’ve got a sub-10 Ironman race under my belt, still chasing that sub-9 dream. While I don't get much time for other sports, I remain enthusiastic and determined—surfing may still happen one day! 🌊

Contact: Want to chat about AI, go for a bike ride, or grab coffee? Send me a DM on X / LinkedIn / Strava.

Popular repositories Loading

  1. runPowerNetworks runPowerNetworks Public

    Jupyter Notebook 12 3

  2. BEP BEP Public

    This repository is devoted to the development of the facial emotion recognition (FER) system as a final bachelor project at the TU/e. Realised by Blazej Manczak. Supervisors: Dr. Laura Astola (Acce…

    Python 5 1

  3. AoM-LineMatching AoM-LineMatching Public

    Detecting, filtering and matching geometrical features from an extraction set to the query image. Project for Superposition in collaboration with Netherlands Institute for Sound and Vision .

    Jupyter Notebook 4 1

  4. ML1Labs ML1Labs Public

    This repo is devoted to the lab exercises of the course 52041MAL6Y realised at the University of Amsterdam, 2020.

    Jupyter Notebook

  5. dl_practicals dl_practicals Public

    This repo is devoted to the pracicals of the course Deep Learning (5204DLFV6Y) realised at the Univeristy of Amsterdam, Fall 2020.

    TeX

  6. former former Public

    Forked from pbloem/former

    Simple transformer implementation from scratch in pytorch.

    Python