About
I'm an AI Engineer and Computer Science researcher building production machine-learning systems, from state-of-the-art computer vision on compute-constrained edge hardware to LLM-powered interfaces over enterprise data. I'm pursuing my M.Sc. in Computer Science with a specialization in Artificial Intelligence at NTNU, after an exchange year at UC Berkeley.
At Astar Consulting I build computer-vision systems for NorgesGruppen Data that run on NVIDIA edge devices, and as a visiting researcher at SINTEF I work on tactile perception for robotic manipulation.
I'm a published author with research across ML systems, recommender systems, LLM robustness, and robotic tactile perception. Previously, I built an LLM-based natural-language interface to Snowflake data at BAMA, and managed sponsor relations for Ascend NTNU.
Programming
AI & Machine Learning
ML Frameworks
Data & Platforms
Systems & Infra
Experience
AI Engineer
Astar Consulting AS · Norway
- Developing a production computer-vision system for the AI Lab at NorgesGruppen Data, the IT arm of Norway's largest grocery retailer.
- Building and training state-of-the-art computer vision models in PyTorch for compute-constrained NVIDIA edge devices.
- Deploying models through optimized DeepStream pipelines, tuning for on-device hardware and latency constraints.
- Operating production systems on the Azure data platform with Azure DevOps CI/CD, from ingestion to serving.
Visiting Researcher
SINTEF · Trondheim, Norway
- Conducting research on tactile perception for robotic manipulation, supervised by Dr. Ekrem Misimi.
- Developing learning-based methods on high-resolution GelSight tactile images, integrated with a Franka Emika robot arm for contact-rich grasping and manipulation.
Data & AI Engineering Intern
BAMA Gruppen AS · Oslo, Norway
- Built a natural-language interface to Snowflake data, letting non-technical business users query internal consumer and market data directly.
- Developed an LLM-based NL-to-SQL pipeline over a curated semantic model, evaluating multiple LLMs to balance accuracy, latency, and usability, and iterating with stakeholders to drive adoption.
- Part of BAMA's group-wide initiative to make data insights more accessible, presented final solution to senior leadership.
Key Account Manager
Ascend NTNU · Trondheim, Norway
- Managed sponsor relations and partnerships for Ascend NTNU, a student organization focused on autonomous drone systems.
- Negotiated and maintained agreements with key industry sponsors, securing funding for technical projects and competitions.
- Led event planning for sponsor showcases and organized financial reporting and budget analysis for the business team.
Research
Trading Engagement for Sustainability: Carbon-Aware Re-ranking for E-commerce Recommendations
2026Syrdal, N.L., Vestrum, A., Bergh, J. · arXiv:2606.04550 (cs.IR) ↗ · UC Berkeley
A carbon-aware re-ranking framework for e-commerce recommendations that trades predicted engagement against inferred product carbon footprint, constructing Pareto frontiers across three product categories on the Amazon Reviews dataset.
Beyond Binary Priorities: Multi-Tier SLA Scheduling for Large Language Model Serving
2026Vestrum, A., Raeesi, A., Roed, H. · arXiv:2608.16336 (cs.AR) ↗ · UC Berkeley
Extended Llumnix's binary priority model to support arbitrary SLA tiers for multi-tenant LLM serving, demonstrating that four priority tiers achieves the best cost-efficiency tradeoff with up to 8.3× prefill and 3.1× end-to-end P99 speedups over INFaaS.
Ask Before You Summarize: NLI-Guided Uncertainty and Clarification-Aware Abstractive Summarization
2026UC Berkeley
A clarification-aware abstractive summarization system that uses NLI-based uncertainty estimation to decide whether to commit to a summary immediately or pause and ask a targeted binary clarification question.
Robustness of In-Context Learning via Curriculum Learning
Oct – Dec 2025UC Berkeley
Studied whether training Transformers on noisy demonstrations improves robustness of in-context learning at inference time, and whether curriculum-based noise schedules provide additional benefits over random noise.
Infrastructure Form Shapes Equitable Access to Essential Services under Flooding in Philippine Cities
2026 · In ProgressUC Berkeley
A reproducible network-accessibility framework measuring how flood-disrupted roads cut off residents from hospitals, schools, markets, and pharmacies across ten Philippine cities, with demographic profiling of the most exposed populations.
Projects
EECS RAG System
2026Python · FAISS · BM25 · sentence-transformers · OpenRouter
Retrieval-Augmented Generation system for answering factoid questions about UC Berkeley EECS, built for CS 288. Combines dense vector search and sparse keyword retrieval with Reciprocal Rank Fusion to maximize answer accuracy.
Prediction of Population Growth
2025Python · scikit-learn · SHAP · Jupyter
Machine learning pipeline predicting country-level population growth rates across 5, 10, and 15-year horizons using socioeconomic indicators from 150+ countries spanning 1960–2016.
Education
University of California, Berkeley
Exchange Student · Coursework in Artificial Intelligence, Computer Science, and Entrepreneurship
Aug 2025 – May 2026
GPA 3.9/4.0
NTNU, Norwegian University of Science and Technology
M.Sc., Computer Science, Artificial Intelligence
2022 – Expected Jun 2027
GPA 4.2/5.0
Certifications
Intro to Snowflake for Devs, Data Scientists, Data Engineers
Coursera
View certificate ↗Beyond Work
Cal Lightweight Rowing
Competed for the University of California, Berkeley lightweight rowing team during my exchange year.
Berkeley Baris Boys
Played for Berkeley Baris Boys Norway in UC Berkeley's IMLeague intramural soccer competition.