Lazar Milikic

Expertise LLMs (Large Language Models), NLP (Natural Language Processing), Automated Reasoning Mission My work focuses on developing robust methods for autoformalization and automated theorem proving using large language models. I aim to enhance the reasoning reliability of modern AI systems by reducing hallucinations, improving predictability, and integrating formal verification techniques and tools into LLM-based pipelines. Before joining EDIC, I completed a Bachelor’s in Mathematics and Computer Science at École Polytechnique (France), graduating summa cum laude and supported by an Excellence Scholarship, followed by a Master’s in Computer Science at EPFL (GPA 5.86/6) where I was the recipient of the EPFL Excellence Fellowship. My previous research spans causal reasoning (ACL 2024), graph neural network–based compiler optimization (CGO 2025), and static program analysis (ICSE 2022). I have also held research roles at Microsoft, Oracle Labs, and Palantir, working on large-scale machine learning systems and compiler optimization technologies. Professionals experiences Research Assistant I developed Vision–Language Distance (VLD), a scalable, multimodal distance-prediction framework for navigation that combines large-scale video pretraining with simulation-trained RL policies. I designed the temporal distance model, its ordinal-consistency evaluation method, and the geometric-overlap noise mechanism that enables reliable deployment by bridging learned perception and robust control. This work introduced a new path toward scalable, goal-conditioned navigation with both image- and language-specified goals.   2025–2025 Robotics Systems Lab, ETHZResearch Assistant At Oracle Labs, I integrated a new GNN-based static profiling model directly into the Graal compiler, replacing the production XGBoost-based profiler and delivering a 4% end-to-end runtime improvement. I developed GraalNN, a two-phase context-sensitive profiling framework that achieves over 10% performance gains on industry-standard benchmarks by leveraging CFG-level graph neural networks. I also built the automated monitoring and retraining pipeline to keep the model aligned with compiler evolution in production environments.   2023–2025 Oracle LabsSoftware Engineer At Palantir, I contributed to both ML and large-scale data infrastructure projects. I extended the SAMv2 architecture to enable user-centered ML for Visual Search to test it against existing internal pipelines. I also implemented aggregation pushdown capabilities in an internal data source, allowing Spark workloads to execute complex aggregations directly at the source and substantially reducing end-to-end query latency. 2024–2024 PalantirResearch Assistant  I developed the first static analysis algorithm tailored for data science notebooks, capable of detecting critical issues such as data leakage and stale-state bugs while preserving interactive performance. I automated large-scale analysis of over two thousand real-world notebooks, demonstrating that our framework could analyze the vast majority within sub-second latency and surface real defects in more than a fifth of tested cells. This work established the foundations for robust notebook correctness checking in production environments. 2021–2022 Microsoft Education Computer Science | Machine Learning, NLP, Causal Reasoning, and Autonomous Navigation 2022 – 2025 EPFL Mathematics and Computer Science | 2019 – 2022 Ecole Polytechnique