MSc Computer Science @ ETH Zürich

Systems, ML infrastructure, and backend engineering.

I build measurement tools and typed AI workflows, with an emphasis on reproducible evaluation and inspectable behavior. Major in Data Management Systems, minor in Machine Learning. Available full-time from summer 2027.

Portfolio · Research & Projects · Blog · LinkedIn

A warm-up selector evaluated on seven public cache traces with a WebLINX-derived synthetic agent overlay.

Key research work:

• Leave-one-trace-out evaluation: 59/63 settings correct (93.7%).
• Mean human-request hit rate of 13.09%, versus 12.94% for the best fixed baseline: +0.141 percentage points.
• Reproduction script checks aggregate inputs, refits held-out folds and validates archived predictions.
• Explicit limits: synthetic overlays, a 64 MiB cache budget and unproven transfer to collected agent traffic.

Repo · Results & Write-up · Technical Note · Slides

Request-level traces and replay tooling for studying the infrastructure behind browser-mediated AI workloads.

Key engineering work:

• Typed request capture, sanitization tools and cache-ready trace exports.
• Public replay inputs: 82,455 scripted request rows and 357,782 LLM-labelled request rows.
• Offline evaluation of six cache policies at five capacities, with request and byte hit rates reported separately.
• A release audit that distinguishes observable CSV contents from collection counts the public snapshot cannot verify.

Repo · Results & Write-up · Replay Data

A TypeScript library for composing AI tasks with Zod schemas, loops, branches and maps.

Key engineering work:

• Runtime schema validation and inspectable execution traces.
• Deterministic mock runtime for testing without inference calls.
• Static pipeline checks and token/cost estimates.
• Reproducible local orchestration benchmarks with raw samples, p50/p95 and clear separation from model latency or quality.

Repo · Engineering Write-up · Benchmark Method

First place in the Chain IQ case challenge, built with a team at START Hack 2026.

A procurement prototype combining language-model parsing with deterministic decision rules and an auditable execution path. The public repository contains the hackathon submission.

Repo · Architecture & Blog Post

Bachelor’s thesis work on traffic/weather integration for vehicle assistance: API adapters, spatio-temporal filtering, TTL caching, fuzzy risk scoring and monitoring.

Repo

• TORCS: behavioral cloning with nearest-neighbor search and KD-tree optimization.
• Broletter: personalized science delivery using arXiv and Telegram. Blog post.

I also contribute to internal infrastructure and backend automation at the ETH Entrepreneur Club.