levineuwirth.org/content/about.md at 42ba2bf972279e520ac6030a33e00bb0e515e9a0

title	tags
Levi Neuwirth — Vita	meta

For a less formal, more detailed introduction to who I am, see Me.

Documents

These are probably what you're looking for. A summary of the key points follows below!

Technical University of Denmark (DTU) — MSc in Computer Science and Engineering. September 2026 – expected 2028. Expecting PhD studies after.
Brown University — Sc.B. in Computer Science and Mathematics. August 2022 – May 2026
DIS Copenhagen / Københavns Universitet — Semester abroad. Fall 2024

My work clusters into three threads:

Machine learning — order-invariant ICD-10-CM embeddings (under review at JAMA Network Open, deployed calculator), the NeuroPose 3D-kinematics system in Liqi Shu's lab at Brown Neurology, and ongoing research engineering at NeuroAI. Undergraduate work has been clinically focused; graduate study broadens the scope.
Computer systems and high-performance computing — the Weenix kernel, a TCP/IP networking stack from scratch in Go, and micro-architectural performance work (SIMD, hardware counters via PAPI, RAPL energy, cross-ISA ports across AVX2 / ARM NEON-SVE / RISC-V V) on Brown's OSCAR HPC cluster.
Reinforcement learning and applied AI — a Magic: The Gathering RL project on OSCAR; production agentic-systems work at xAI on grok-code-fast-1; reasoning, evaluation, and red-teaming contracts with Anthropic, Mistral, and OpenAI.

Computer vision and security thread through all three but do not stand on their own.

Shu L, Neuwirth L†, Wang X†, Zheng H†. Beyond Comorbidity Indices: An Order-Invariant ICD-10-CM Embedding for Readmission and Mortality Prediction. Under review at JAMA Network Open, 2026. Preprint · Calculator · Code
Neuwirth L. Where Does SIMD Help Post-Quantum Cryptography? A Micro-Architectural Study of ML-KEM on x86 AVX2. Technical report, Brown University Department of Computer Science, April 2026. Report · Artifact

NeuroPose clinical-implications manuscript. In preparation; target submission 2026–2027.
SIMD / PQC Phase 2 & Phase 3. Hardware performance counters (PAPI), RAPL energy, and cross-ISA ports (ARM NEON/SVE, RISC-V V).
Semantic-embeddings citation project. Early-stage work with NeuroAI, preprint expected summer 2026.
Magic: The Gathering reinforcement learning project. Early-stage work through Brown's HPC, expected late 2026.

Early Detection of Neurological Disorders through Video-Captured Kinematic Analysis. Ma J, Arms S, Kaneira L, Lall M, Chen K, Cabral W, Man D, Neuwirth L, Shu L. Poster, Brown / Rhode Island Hospital Neurology Summer UTRA Symposium, August 2025.
**"Order-Invariant ICD-10-CM Embedding for Readmission and Mortality Prediction: Toward Multimodal Generative Patient Models"**Shu L, Neuwirth L†, Wang X†, Zheng H†. IEEE/ACM Conference on Connected Health: Applications, Systems and Engineering Technologies (CHASE), August 2026, in review.

†Equal-contribution undergraduate authors.

xAI Summer 2025, remote. Contributed to training of grok-code-fast-1. Integrated LLMs into autonomous agent frameworks spanning 20 tools across thousands of production workflows and codebases; diagnosed and resolved 50+ agentic failures, improving tool-execution rates by >15%, reducing fatal tool-usage errors by >40%, and reducing API usage by >20%. Contributed to safety systems, reducing hallucination rates by >50% in controlled agent-behavior tests.
NeuroAI Present. neuroai.health. Early-stage venture of academics and clinicians integrating deep learning, reinforcement learning, and generative AI into clinical and research workflows; leading research-engineering across model development, deployment infrastructure, and system design.
Shu Laboratory, Brown Department of Neurology — Undergraduate Researcher and Technical Lead. October 2023 – Present. Technical lead on NeuroPose; co-lead developer on the ICD-10-CM embedding model.
Independent Research Contracting — Anthropic, Mistral, OpenAI. 2025 – Present Expert reasoning contributions in code and mathematics for agentic workflows, agentic task design and evaluation, AI safety, and red-teaming.

Weenix — Unix-like kernel in ~7,000 lines of C: virtual memory, VFS, system calls, threading, device drivers, interrupt handlers, and file systems. Custom linker support for running userspace x86-64 ELF binaries; extended with pipes and userspace preemption.
Networking Stack from Scratch — TCP/IP, RIP, UDP, and DNS in Go, supporting file transmission of up to 1 GB across 8-node networks. Extended with a fully RFC-compliant SSH implementation (2,000+ additional lines) supporting sustained sessions of arbitrary length.
SIMD / PQC Performance Study — Hand-written AVX2 assembly for ML-KEM / Kyber. 35×–56× speedup over compiler-optimized C for core NTT arithmetic; 5.4×–7.1× end-to-end KEM speedup, with a full statistical-analysis pipeline on Brown's OSCAR cluster.

For the complete index — additional artifacts, deployed ML, and smaller tools — see /cv/projects/.