Amund Tveit
Trondheim, Norway

Amund Tveit

Principal AI Engineer · PhD in Computer Science · Trondheim, Norway

I am a computer scientist and AI engineer based in Trondheim, Norway. Work-wise, I focus on innovation and efficiency with AI and agentic engineering for Bank Tech. My technical interests span model training and tuning, large-scale verifiable agentic engineering, and the intersection of compilers/interpreters, formal methods and AI.

Since the 1990s I have worked across academia, startups, and global technology companies: starting with RS/6000-AIX systems at IBM Canada and financial pricing software at Storebrand, earning a PhD in Computer Science at NTNU, leading engineering teams at Google and YouTube in Trondheim and Zürich, founding Memkite and Atbrox, serving as VP of Data at Zedge, working as Principal Data & Applied Scientist and Principal Product Manager at Microsoft, and now serving as Principal AI Engineer at Tieto.

1990s – Present

Career Timeline

2026 – Present

Principal AI Engineer

Tieto · Trondheim, Norway

Driving innovation and efficiency with AI and agentic engineering for Bank Tech.

2025 – 2026

AI Engineer

Atsentia · Trondheim, Norway

Built high-performance open-source Mojo systems libraries (mojo-json, mojo-vec, mojo-tokenizer, mojo-contrib) and Apple Silicon inference tooling.

2024 – 2025

Principal Product Manager

Microsoft · Trondheim, Norway

Product Manager for latency and quality evaluation on M365 Copilot Search. Advisor and AI Evaluation & Measurement Area Chair for Microsoft MLADS+. Published arXiv:2504.16041 on Muon optimizer accelerating grokking.

2019 – 2024

Principal Data & Applied Scientist

Microsoft · Trondheim, Norway

Spearheaded synthetic data workflows for agility and advised engineering teams on model inference, fine-tuning, on-device LLMs, and compliant data science tooling.

2018 – 2019

Engineering Manager, YouTube

Google · Zürich, Switzerland

Managed software engineers on YouTube Channel Memberships across billion+ MAU surfaces (homepage, watch page, channel, live chat, YouTube Studio) and led an AI innovation feature with Google AI.

2016 – 2018

Vice President of Data

Zedge · Trondheim, Norway

Led data science, analytics, search, and recommender systems for 30M+ MAU. Replaced Hadoop and legacy search with ClickHouse and Vespa on Google Cloud; shipped deep learning audio/image similarity, real-time recommendations, and image super-resolution.

2013 – 2015

Founder & CEO

Memkite · Trondheim, Norway

Created open-source DeepLearningKit (Convolutional Neural Networks in Swift & Metal for iOS), built the first App Store app in its category, and patented brain-signal-based instant search for consumer EEG.

2009 – 2015

Co-Founder

Atbrox · Trondheim, Norway

R&D consulting from startups to Fortune 500 companies, research in Hadoop/MapReduce, GPU computing, cloud, and IoT, plus early-stage Nordic and US edtech investments.

2008 – 2009

Engineering Manager

Google · Trondheim, Norway

Architect and tech lead for large-scale production monitoring infrastructure; co-organizer of the Google 2009 EMEA Faculty Summit on Software Engineering for the Cloud.

2005 – 2008

Software Engineer

Google · Trondheim, Norway

Tech lead and developer for Mobile Google News for iPhone and iPod Touch; intern manager for Google Translate (NO/SE/DK), Google News Scandinavia, and large-scale data visualization.

1999 – 2005

PostDoc & PhD in Computer Science

NTNU · Trondheim, Norway

PostDoc in medical/health informatics (anonymization, text mining, co-advising PhD and MSc students). PhD thesis (1999–2004): Customizing Cyberspace: Methods for User Representation and Prediction (850+ citations on Google Scholar; h-index 14). Also founded Agentus (1999) alongside the ElComAg intelligent-agents research program.

1998 – 1999

Software Engineer & Intranet Project Lead

Storebrand ASA · Norway

Built software for pricing life insurance products and statistical prognosis models for mutual funds and savings products, developed and tested a Lotus Notes insurance sales system, and led intranet initiatives.

1996 – 1997

Software Engineer, Call Centre Solutions

IBM Canada · Canada

RS/6000-AIX systems engineering covering Java, C++ GUI development, C asynchronous communications, Oracle (Pro*C and PL/SQL), DirectTalk/6000 state tables and custom servers, Perl scripting, and AIX administration.

1990 – 1997 (Summers)

Summer Internships & Early Engineering

Elkem, Tallgrass, Open Systems Consultants & Viewtech · Norway

Summer jobs and internships from industrial process control at Elkem (FeSi furnace control room, DataEase, FixDMACS, and an ARMAX model in C at Elkem Research) to early WWW tooling at Tallgrass Technologies, Unix porting in Perl/ksh at Open Systems Consultants, and C/Motif at Viewtech.

Intellectual Property

Issued US Patents

Full portfolio on Google Patents
PatentYearCompanyTitle
US 12,670,198 B2 2026MicrosoftTextual summaries in information systems based on personalized prior knowledge
US 12,664,485 B2 2026MicrosoftMeasuring probability of influence using multi-dimensional statistics on deep learning embeddings
US 11,669,860 B2 2023GoogleMethods, systems, and media for automated compliance determination of content items
US 11,636,138 B1 2023GoogleTemporal layers for presenting personalization markers on imagery
US 9,737,231 B2 2017MemkiteBrain signal-based instant search
US 8,600,802 B1 2013GoogleAdvertisement marketplace
US 8,584,013 B1 2013GoogleTemporal layers for presenting personalization markers on imagery
US 8,489,614 B2 2013GoogleRanking academic event related search results using event member metrics
US 8,392,433 B2 2013AtbroxSelf-indexer and self indexing system
Academic & Preprints

Selected Research & Publications

All papers & 16 PDFs
arXiv:2504.16041 · Apr 2025

Muon Optimizer Accelerates Grokking

Empirical evaluation across seven modular arithmetic tasks comparing Muon vs. AdamW and softmax variants (standard softmax, stablemax, sparsemax). Muon reduced the mean grokking epoch from 153.09 to 102.89 (t = 5.0175, p = 6.33e-08), accelerating the transition from memorization to generalization.

Code & Benchmarks

Selected Open Source

Full software catalog
Mojo · AI Agents

pi-mojo

A pure Mojo port of the PI AI Agent Toolkit for fast, low-overhead agent workflows.

Deep Learning · Theory

one-op

Deep learning and functional approximation experiments built on the Exp-minus-Log (EML) Sheffer primitive.

Swift · Metal

DeepLearningKit

Open-source (Apache 2.0) GPU-accelerated deep learning framework for Apple iOS, OS X, and tvOS written in Swift and Metal (2015–2016).

Education & Service

Selected Credentials

Udacity2017

Deep Learning Nanodegree

Projects covering Convolutional Neural Networks, Recurrent Neural Networks (LSTM/seq2seq), and Generative Adversarial Networks (GAN).

Contact

Best way to reach me: LinkedIn for professional or speaking inquiries, X (@atveit) for technical discussions, or GitHub for code and benchmarks.

amund.blog publishes personal views, notes, and independent research experiments. Looking for earlier writing? Browse the 2008–2025 archive