Mihai-George Licu

• Designed hardware migration predictors for CXL-based memory systems that decide when data is worth moving to local DRAM, reducing the performance cost of remote memory.
• Cut the runtime of a C++ trace-driven simulator for out-of-order multi-socket CPUs by 2x and its memory usage by 40%, and improved the accuracy of its timing and logic models.
• Managed a 300-CPU Slurm cluster the simulations run on, including building a dashboard for its administration and logging.

Swiss Federal Institute of Technology

• Generalised the ingestion layer of a Python population synthesis data pipeline onto new statistical and geospatial sources, extending it to new geographical regions for MATSim public transit simulations.
• Enhanced the pipeline's population fitting and location assignment algorithms, improving how closely the synthetic data matched observed demographics and travel behaviour.

European Organization for Nuclear Research

• Upgraded CERN's Wappalyzer-based scanner from Python 2 to Python 3, added detection features, and integrated Kerberos SSO to reach 1000+ previously unscannable applications.
• Evaluated unified YubiKey SSH authentication across CERN's Linux, macOS, and Windows estate, prototyping an integration and documenting FIDO2 protocol limits and per-platform requirements.

Multinational investment bank

• Built a CI/CD pipeline on Google Cloud Platform for an internal Spring Boot and React training platform serving 1500+ employees, automating its builds, tests, and deployments.
• Engineered the platform's notification microservice on Google Cloud Functions, delivering event-driven email and SMS, and wrote the scripts behind its usage analytics.

Swiss Federal Institute of Technology

• Built the Docker-based lab environment and exercises for a graduate Systems Security course taken by 200+ students, covering symbolic execution (angr, KLEE), taint analysis (Triton, DECAF), and sandboxing.
• Benchmarked a research WebAssembly sandbox that uses Intel Memory Protection Keys for isolation, and debugged the crashes those runs exposed.

C++, Python

• Optimised the attention layer of a geometric algebra transformer with online softmax, fast exponent approximation and FlashAttention tiling, averaging 3.8x faster inference than the PyTorch reference while executing 75% fewer operations.

Rust

• Wrote a cycle-accurate, out-of-order superscalar processor simulator in Rust executing a subset of RISC-V, modelling register renaming, a 32-entry reservation station, and precise exception handling.