Software Engineer: Keystone Project (Machine-Verified LLM Inference) — EPFL

Vollständige Beschreibung

Mission

The Keystone project, an ARIA-funded collaboration between EPFL and Imperial College London, aims to build a formally-verified ML inference engine, demonstrating that AI can help make verified systems competitive with unverified systems in terms of development effort, features, and performance. Its missions include the design and implementation of verified inference components, the formalization of GPU kernel semantics, the development of AI-assisted proof engineering workflows, and the open-source release of specifications, proofs, and verified artefacts.

We seek an excellent software engineer to play a central role in turning verified research prototypes into a production-grade, high-performance inference engine. Prior experience with formal verification is welcome but not required: a strong systems engineer with the motivation to learn proof-assistant technology will thrive in this role.

Main duties and responsibilities

• Bring technical expertise in systems programming and performance engineering to support the project’s research; collaborate with the research teams at EPFL and Imperial College London • Design, implement, and maintain core components of the verified LLM inference engine, including the runtime and glue code connecting extracted verified code, GPU kernels, and drivers • Organize and manage the project’s engineering infrastructure: differential testing against reference engines (vLLM, SGLang), continuous integration for code and proofs, and performance benchmarking • Develop and maintain agentic AI pipelines for specification autoformalization, proof generation, and proof repair • Write documentation, procedures, and recommendations to ensure reproducibility of the project’s artefacts • Diagnose, prevent, and repair failures and regressions across the software stack • Analyze the security level and trusted computing base of the components we develop, and contribute to red/blue team exercises within the ARIA programme • Contribute to open-source releases and engage with their user communities

Profile

• Higher degree in computer science or education deemed equivalent; experience in the field • Excellent technical knowledge of systems programming, and strong programming ability in several of: Python, C/C++, Rust, OCaml, or other functional languages • Knowledge of one or more of the following, with strong motivation to grow in the others: • GPU programming (CUDA, Triton, PTX) or high-performance computing • ML inference or serving systems (vLLM, SGLang, PyTorch internals, or similar) • Interactive theorem proving (Rocq, Lean, HOL, Isabelle, or similar) or other formal methods • Experience maintaining development tooling: build systems, continuous integration, and test infrastructure • Experience using LLM-based development tools or building agentic workflows is a plus • Mastery of English indispensable (oral and written); French is an asset but not required • Sense of priorities, integrity, and autonomy in your work • Team spirit and aptitude for conducting technical investigations and implementations; excellent ability to communicate with varied audiences, from proof engineers to systems researchers • Strong sense of service and spirit of initiative

We offer

• The possibility to join a dynamic and stimulating team on a high-profile ARIA-funded project • A multicultural and academic working environment of high quality • Opportunities for continuing education and professional development • Excellent working conditions • Generous access to frontier AI models and high-performance compute • Funded travel for collaboration between Lausanne and London, and for conferences

Informations

Only applications submitted through the online platform are considered. You are asked to supply: • A brief cover letter (pdf, up to 2 pages).

And in one PDF

• A CV, including links to open-source contributions or representative projects where applicable. • Contact details for 3 refere

Fragen zu dieser Stellenanzeige