Skip to content
View levineuwirth's full-sized avatar
💭
https://git.levineuwirth.org
💭
https://git.levineuwirth.org

Highlights

  • Pro

Block or report levineuwirth

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
levineuwirth/README.md

Levi Neuwirth

Work · CV · Website · Email · GPG · ORCID

I work on technical AI assurance, establishing verifiable claims about what a model actually did, when the operator, the evaluator, or the model itself may not be trusted. This spans cryptography, systems, evaluations, and mathematics, precisely the shape of my background. I was a MARS V fellow with the Cambridge AI Safety Hub, mentored by James Petrie (FLI); before that, I studied mathematics and computer science at Brown University.

I am open to full-time research and research-engineering positions worldwide.

Selected work

Verifiable LLM inference (MARS V, ongoing) — zero-knowledge proofs of LLM inference that mutually distrustful parties can check on their own hardware. Anthropic's FlashProve post calls VerInf "the only earlier proof of Llama 4 Maverick that we know of." My work, merged upstream: a dry-run profiler and cost model validated at full Llama 4 Maverick scale; a weight split whose partitioned proofs are byte-identical to ordinary ones; weight caches; the integration and hardening of the expert-weight bridge, which I now develop; and repairs to the attention, activation, and surprisal constraints after a collaborator's review, each counterexample now a regression test. Top-k routing, the form Kimi K2 uses, runs on a toy model (merged PRs · essay). Paper expected Q4 2026.

Frontier-model evaluation and red-teaming — reasoning evaluation, red-teaming, RL environments, and evaluation research under contract. What I can discuss publicly is limited.

proof-broker — a certificate boundary for untrusted proof search. Lean 4 and Rocq goals go through a shared IR to SMT solvers, saturation provers, and language models; a result counts only if its certificate checks and the home kernel accepts the proof. R4 closed 19/19 targeted obligations in an unmodified VerInf file. R6 put a live model behind the boundary under an audited protocol: all 80 feasible certificates verified, yet proofs covered 6 of 15 obligations. Reconstruction, not proposal, was the constraint. (Write-up)

icd_embeddings — order-invariant ICD-10-CM embeddings for 30-day readmission and post-discharge mortality, trained on the Nationwide Readmissions Database. Our manuscript is under review at JAMIA.

Also

Systems and performance: pmacs (Rust-cored, Lua-scripted editor), kyber-simd-profiling (SIMD post-quantum cryptography, cross-ISA), LeVCS (federated version control with signed authority chains), the Weenix kernel, and a TCP/IP stack written from scratch. Mathematics: two graph-theory preprints, on tree-ball geometry and growing-radius domination. Research engineering: neuropose, 3D kinematics in Liqi Shu's lab at Brown Neurology.

Primary Tools

Rust, Python, C/C++, OCaml, Haskell, Go. Cryptography, distributed systems and HPC, formal methods, ML evaluation.

I have worked with many other tools in the past to various extents, and the above list is certainly not exhaustive.


Most of my development is being moved to my Forgejo instance, mirrored here. Current is what I'm actually working on this month, while Vita is the most formal record.

Pinned Loading

  1. JamesPetrie/VerInf JamesPetrie/VerInf Public

    Python 18 5

  2. pmacs pmacs Public

    Parallel Emacs — a Rust-cored, Lua-scripted editor in the Emacs tradition. Development, issues, and PRs happen here; self-hosted mirror at git.levineuwirth.org.

    Rust

  3. icd_embeddings icd_embeddings Public

    A permutation-invariant deep-learning model for 30-day unplanned readmission and 30-day post-discharge mortality, trained on ICD-10-CM diagnosis-code sets from the Nationwide Readmissions Database.…

    Python

  4. kyber-simd-profiling kyber-simd-profiling Public

    Profiling ML-KEM (Kyber): performance and energy implications of SIMD-optimized post-quantum cryptography.

    Go

  5. neuropose neuropose Public

    A 3D pose-estimation and kinematic-analysis system for neurological-recovery research, developed in Liqi Shu's cross-disciplinary laboratory at Brown University.

    Python 1

  6. arcana arcana Public

    A high-performance Magic: The Gathering rules engine in Rust, designed as a substrate for reinforcement-learning research.

    Rust 15