Work — Frederik Willemsen

01 Civic technology · 48h hackathon · TUM.ai Agora Hacks

agorix.eu

An AI civic copilot that turns Munich's 30+ municipal sites into one conversation.

Built in 48 hours by a team of three at TUM.ai Agora Hacks. Demo'd live to a panel of leading professors in politics and computer science, including IBM's Head of AI. Live since the hackathon weekend.

From kickoff to live demo. 30+ municipal sources integrated into one chat interface.

Problem

Munich's municipal information lives across 30+ websites with inconsistent search, opening hours, and registration flows. Newcomers — especially international students — spend hours stitching information together.

Solution

An LLM-routed query layer over the city's real-time data. Free-text questions in, citizen-friendly answers with sources and registration links out.

TypeScriptPythonNext.jsLLM · RAG

What I built

A working conversational prototype: free-text Munich questions in, natural-language answers out.
LLM-driven Q&A wired to a starter set of municipal data sources — a proof-of-concept for the broader vision below.
Source-attribution pattern integrated into responses, so every answer points back to where it came from.
Built, polished, and deployed live at agorix.eu in time for the judging demo.

Achieved

Demo'd live to a judging panel of leading German professors in politics and computer science, including IBM's Head of AI.
Live and reachable at agorix.eu through to today.
Three-person team, 48 hours, end-to-end prototype shipped on time.

The vision

LLM router classifying each query by district + topic before retrieval.
Crawler + normalizer pulling structured data from 30+ Munich gov sites into a single schema.
End-to-end citation pipeline — every answer linked back to the exact municipal page.

0hBuild time

0×Person team

0+Data sources integrated

Visit agorix.euSource on GitHub

02 Practical Course · TUM · 4-month project · 3-person team

Molecular dynamics framework

A high-performance MD simulator in C++. Best in cohort by 60%.

Eight teams, one final benchmark, four months. Three of us built a particle-based MD simulator from scratch — and ended up sixty percent ahead of the next-best team.

Faster than the next-best team on the end-of-course benchmark. 1 of 8 teams in the cohort.

Try it

Below: a live 2D Lennard-Jones simulation running the same linked-cell algorithm that drove the cohort win, ported from C++ to JavaScript and rendered in canvas. Click anywhere to seed a particle, drag any particle to perturb the system, scrub the slider to heat or freeze it.

Live Lennard-Jones · linked-cell · 2D

Click to add · drag a particle · scrub the temperature

TemperatureT = 1.00

0 particles

What we built

A particle-based MD simulator implementing linked-cell neighborhood search, classical force fields, periodic boundaries, thermostatting, and trajectory I/O — parallelized with OpenMP, designed for cache friendliness end-to-end.

C++OpenMPCMakeHPC

Why it won

We profiled from week one. Picked the right tradeoffs (memory layout > algorithmic cleverness for this workload). Refused to add features that weren't in the hot path. Tested every commit against a benchmark harness we built ourselves.

Built into the simulator

Linked-cell neighborhood search — the single biggest perf win vs. the cohort.
Lennard-Jones + harmonic bond potentials, with cutoff handling.
Periodic boundary conditions with image cells.
Berendsen + velocity-scaling thermostats for equilibration.
VTK + XYZ trajectory I/O for visualization in ParaView.
Structure-of-arrays particle layout — cache-friendly inner loops.
OpenMP-parallel force evaluation, with thread-local accumulators.

Achieved

Best overall performance — 60% faster than the next-best team.
1 of 8 teams in TUM's Scientific Computing (PSE) practical.
Three-person team, led end-to-end including architecture decisions and perf reviews.
Open-sourced on GitHub as MolSim.

+0%Faster than runner-up

0Competing teams

0×Person team, led

Source on GitHub (MolSim)

03 Research · BSc thesis · TUM PEML Chair · 6 months

ML-tuned HPC simulations

Data-driven configuration selection for particle simulations on CoolMuc-4. Open-sourced in AutoPas.

Can ML pick algorithmic configurations for HPC particle simulations better than hand-tuning? It can — by sixteen percent on CoolMuc-4. Six months at TUM's Chair of Physics-Enhanced Machine Learning. Thesis grade 1.7.

Runtime speedup over hand-tuned baselines on CoolMuc-4. Generalizes across grid geometries — no per-grid re-tuning.

Method

Random forests on execution traces predict per-grid optimal configurations. K-means + PCA on grid features identify which grids share a tuning regime. Performance-driven experimentation closes the loop.

C++PythonPandasscikit-learn

Result

16% runtime speedup over legacy hand-tuned approaches on a large-scale HPC system. Generalizes across grid geometries — no per-grid re-tuning. Open-sourced as a patch to AutoPas.

Built

Trace logger capturing per-configuration execution metrics on CoolMuc-4.
Feature pipeline (Pandas) extracting grid geometry + workload features.
Random-forest classifier predicting optimal config per grid.
K-means + PCA clustering to identify grids that share a tuning regime.
Closed-loop tuner driving performance-driven experimentation.
Integration patch upstreamed to AutoPas — TUM's open-source HPC particle library.