Postdoc: Keystone Project (Machine-Verified LLM Inference)

EPFL; Lausanne; Vor Ort; CHF 110.000 - 140.000; Vollzeit

Benefits dieser Stelle

International working environment

Excellent working conditions

Frontier AI models access

Travel for collaboration

Zusammenfassung

The Keystone project at EPFL in Lausanne seeks exceptional candidates in formal methods and computer systems to advance machine-verified ML inference engines. You will contribute to GPU semantics, the coordination layer, and AI-assisted proof workflows, collaborating with Imperial College London on a high-profile ARIA effort.

Applicants should hold a PhD (or near completion) in CS with strong publication records and proficiency in Python, C++, OCaml, and related tools.

Qualifikationen

• PhD (or nearing completion) in computer science or closely related field.
• Background in formal verification, programming languages, systems, or ML.
• Research experience in interactive theorem proving (Rocq, Lean, HOL, Isabelle, or similar), GPU programming/semantics, compilers, concurrency, or ML systems.
• Strong software engineering across multiple languages (e.g., Python, C++, OCaml).
• Publication record relative to career stage in international venues.

Aufgaben

• Conduct research related to the Keystone project, including designing and building a machine-verified LLM inference engine.
• Explore formalizing GPU kernel semantics in a proof assistant and verify high-performance inference kernels.
• Develop and verify the inference coordination layer (batching, KV-cache, scheduling) with extraction to executable code.
• Develop agentic AI workflows for specification autoformalization, proof generation, and proof repair.
• Engage with ARIA programme through red/blue team exercises and sprint reviews.

Kenntnisse

Formal verification

Programming languages

Python

C++

OCaml

Lean/Rocq

AI/ML systems

Ausbildung

PhD in CS

Tools

Rocq

Lean

HOL

Isabelle

Jobbeschreibung

Mission

The Keystone project seeks to advance knowledge in the following domains:

• Formal verification and interactive theorem proving (Rocq, Lean)
• Secure and high-performance computer systems, including ML infrastructure

We seek outstanding candidates working in formal methods and/or computer systems with interests in one or more of the following areas: machine-checked verification of systems software, GPU kernel semantics and verification, and AI-assisted proof engineering.

The position is part of an ARIA-funded collaboration between EPFL and Imperial College London whose goal is 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.

Main Duties And Responsibilities

• Conducting research related to the Keystone project, including contributing to the design and construction of a machine-verified LLM inference engine. The precise focus will be discussed with the successful candidate depending on their background, expertise and affinities. Possible directions include:
  • Formalizing GPU kernel semantics (e.g., PTX/Triton) in a proof assistant and verifying high-performance inference kernels (numerical accuracy, memory safety, data race freedom, functional correctness)
  • Implementing and verifying the inference coordination layer (batching, KV-cache management, scheduling) in Rocq/Lean, with extraction to executable code
  • Developing agentic AI workflows for specification autoformalization, proof generation, and proof repair
• Build a strong network in the fields of formal verification, systems, and ML infrastructure, including close collaboration with the Imperial College London team (regular visits to London are expected)
• Contribute to open-source releases of specifications, proofs, and verified artefacts
• Engage with the ARIA programme, including red/blue team exercises and sprint reviews

Profile

• PhD (or nearing completion of) in computer science or a closely related field
• Background in formal verification, programming languages, systems, or ML
• Research experience in one or more of: interactive theorem proving (Rocq, Lean, HOL, Isabelle, or similar), GPU programming or semantics, compilers, concurrency, or ML systems (e.g., vLLM, SGLang, or similar inference engines)
• Strong computational and analytical skills, including solid software engineering ability across multiple languages (e.g., Python, C++, OCaml, functional languages)
• Experience using or evaluating LLM-based tools for code or proof generation is a plus
• Publication record (relative to your career stage) in internationally leading journals and conferences
• Independent, creative, and solution-oriented
• Excellent written and oral communication skills in English
• Strong motivation to explore new research domains at the intersection of AI and formal methods
• Good team spirit and enthusiasm for working in a distributed, multi-institution team

We offer

• A stimulating and international working environment
• Excellent working conditions
• Opportunity to perform state-of-the‑art research in one of the most dynamic scientific institutions in Europe, within a high‑profile ARIA-funded project
• Opportunity to interact with internationally renowned experts at EPFL and Imperial College London, and with a strong team of postdoctoral researchers and PhD students
• Generous access to frontier AI models and high‑performance compute for proof‑assistant workloads
• 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 with a publication list.
• A research statement (pdf, up to 3 pages).
• Contact details for 3 referees.

For any further information, please contact: Nate Foster ([e-mail]).

More information can be found on https://laser.epfl.ch/.

Contract Start Date : 01.12.2026, or to be determined

Activity Rate : 100.00

Contract Type: CDD

Duration: 1 year, renewable (project duration permitting)

Reference: 2476