projects
Hackathons
Verifiable Benchmarks
Privacy AI Blockchain
I was glad to participate in the EigenLayer Hacker House during Berlin Blockchain Week 2025. Thanks to EigenLayer for providing us with co-working space and accommodation for the duration of the Blockchain Week. I had the opportunity to build, meet great people, and attend numerous events throughout Berlin.
I collaborated on a project with my friend Alexander Semenov from the TUM Blockchain Club. We built upon work we had previously developed at ETHGlobal in Prague. Our earlier project, which computed SNARK proofs of inference, was quite slow in execution. This time, we explored a different approach using the TOPLOC, which is implemented as a library. This approach provides weaker guarantees than SNARK proofs but is significantly faster and more practical for real-world applications. We integrated the entire system with the EigenLayer ecosystem, enabling proofs to be published on-chain with verification handled by an AVS (Actively Validated Service).
Proofs of inference
🥇 1st Place Hedera 🥇 1st Place Protocol Labs
Privacy AI zkSNARK Blockchain
AI systems with a client-server architecture often function as black boxes, meaning users cannot independently verify whether a prediction was generated correctly using the intended model. To address this, we developed a protocol that allows users to verify the correctness of AI model predictions. For this, we utilized the established EZKL library, which compiles models in ONNX format into arithmetic circuits. These circuits can then be evaluated and proven as SNARK proofs. Specifically, we demonstrated inference verification using the Halo2 proving system.
Next, we built an application around this protocol. We created a web interface where users provide input to a model and receive both the output and a corresponding proof. This proof is then stored on IPFS. To enable on-chain verification, we implemented a smart contract verifier deployed on Hedera.
CryptoCaching
🥉 3rd Place Hedera
Tokenization P2P Blockchain
A blockchain-based alternative to traditional GeoCaching. We solved the admin maintenance problem by creating a P2P network to claim and verify cache discoveries. Each cache contains an NFC tag reprogrammed by users, with a token passing mechanism implemented on Hedera to prevent false claims of unvisited caches.
Misc
Organizing Ethereum Research Challenge
At the conference we had Ethereum Foundaiton as one of the main sponsors. I had a chance to look at hackathon from different side as an organizer. I was given full freedom in structuring the event. The theme of the event was MEV, and rather than providing participants with rigid tasks, I wanted to create an open-ended challenge where participants could work on topics that interested them. Twitter status.
In this 24-hour event with $5,000 prize pool, we received some fantastic submissions from experienced teams while also successfully onboarding newcomers. I’m really glad I had the chance to organize this event and learned a lot about what it takes to run a hackathon.
Elliptic Curve Arithmetic
Rust Elliptic curves Cryptography
EC arithmetic based on blog: Elliptic Curve Cryptography: A Gentle Introduction
A Rust implementation of elliptic curve cryptography primitives that includes efficient point addition, scalar multiplication and multiscalar multiplication. The repository features different scalar multiplication methods and implements Pippenger's algorithm for optimized multi-scalar multiplication. All operations are implemented over finite fields.
University Projects
Optimization of fluid dynamics simulation
C++ High-Performance-Computation
This final project was part of the High-Performance Computing Praktikum at TUM. We worked on implementing a parallelized tsunami simulation, with a primary focus on single-core parallelism using SIMD operations. Our approach leveraged x86 architecture features such as vectorized addition and multiplication. Additionally, we addressed load imbalances present in the existing code to improve overall performance.
Efficient Möbius Computations on Multipermutations
Rust Combinatorics Algorithms
This Rust project efficiently computes the Möbius function for multipermutations. We model multipermutations within a poset to capture their ordering and interval structure, which are crucial for the recursive algorithm with memoization we employ. The recursive nature of the Möbius function makes it computationally challenging, so our focus is on achieving high efficiency. The Möbius function offers insights into permutation patterns which is a research interest of advisor VÃt JelÃnek.
Classification of magnetic phases by graph neural networks
PyTorch GNN Physics
This project, supported by the Student Faculty Grant, explores the application of graph neural networks (GNNs) in physics. We encoded the magnetic configurations of the Ising model as graphs, where nodes represent spins and edges capture interactions. Using this model, we aimed to predict configurations with minimal energy. This problem is particularly interesting as it is known to be NP-complete. We experimented with various GNN architectures using PyTorch and concluded with a report on our results, highlighting the most effective approach.