Selected work where curiosity turned into measurement: find the structure, build the experiment, and make the result clear enough that someone else can continue it.
Semester project (Communications and Signal Processing Group, Imperial) under Prof. Cong Ling, spring 2026—ongoing: implement the Macaulay-matrix approach to compute linear-strand Betti numbers of code ideals, run reproducible experiments on random versus structured code families, and analyse when alternant and Goppa codes deviate measurably from random-code behaviour for the Syzygy-based distinguisher.
Imperial (winter 2025): leaky integrate-and-fire spiking neural network trained with surrogate gradients to decode 2D cursor velocity from monkey motor-cortex spike trains, compared against a non-spiking baseline.
ETH Translational Neuromodeling Unit (spring 2025), Prof. Klaas Enno Stephan: test whether baseline resting-state fMRI predicts response to a single ketamine infusion in major depressive disorder (NIMH open data; after QC, n=26). Spectral DCMs at 4 / 11 / 15 nodes (21-node too costly); A-matrices feed logistic regression and related models for binary/multi-class response and MADRS regression. Best: 4-node DMN + logistic regression (about 73% accuracy, F1 0.71, macro-recall 0.82); larger DCMs near 65%. Finer stratification and continuous outcomes weak on this cohort—parsimony beats complexity; future validation and clinical covariates.
ETH Automatic Control Laboratory (spring 2025), Prof. Florian Dörfler: multi-agent informative path planning for online graph exploration with node-wise beam search, adaptive Voronoi partitioning, and a bitmap-based hidden-path mechanism to limit redundant exploration. Compared Euclidean Voronoi, Dijkstra–Voronoi, decentralised hidden-claim, and spectral room assignment; Voronoi coordination balanced quality and cost, with Dijkstra–Voronoi NBS+ strongest on mean gain at similar runtime; spectral layouts cleaner but less flexible.
Bachelor thesis (fall 2024), Trapped Ion Quantum Information Group, Profs. Lukas Novotny and Jonathan Home: integrated a network-analyzer module into BLOOD’s Red Pitaya FPGA, C++ server, and multi-client stack so laser feedback loops can be characterised without external gear. Validated PyRPL-style analysis against Analog Discovery 2, implemented sweep/I/Q flows with multi-client safety, matched Bode plots to PyRPL—closed-loop transfer-function measurement inside BLOOD for leaner lab workflows. Listed on TIQI’s public semester-theses page.
ETH 3D Vision (2025): Surf-LEAF converts urban Gaussian-splatting reconstructions into analysis-ready meshes—surface sampling from splats, advancing-front reconstruction, post-processing for holes and artefacts. On TartanAir, ~10% lower Chamfer than strongest baseline, sharper than Poisson/TSDF-style methods; qualitative gains on real outdoor scenes for digital twins, navigation, and urban modelling.
ETH (spring 2024): developed and tested an FPGA-based control and measurement framework for superconducting-qubit experiments.