Morten Svendgård

"Polynomial Regression as a Task for Understanding In-Context Learning Through Finetuning and Alignment"

• Proposed polynomial regression as a structured function class for studying prompting, alignment, and in-context learning in transformers.
• Ran large-scale experiments on Berkeley compute clusters.
• Worked under Prof. Anant Sahai.

• Graduate-level coursework in deep learning, robotics, and autonomy — including large-scale training runs on university GPU clusters.
• Regularly turned recent research papers into working prototypes and experiment code.

Selected courses:

• EECS C206B — Robotic Manipulation and Interaction (Shankar Sastry)
• MECENG 292B — AI for Autonomy (Wei Zhan, Yuxin Chen)
• COMPSCI 282A — Designing, Visualizing and Understanding Deep Neural Networks (Anant Sahai)
• ELENG 290 — InContext: Understanding in-context learning in language models via simple function classes (Anant Sahai)

• Control theory, signal processing, machine learning, and robotic systems.
• Thesis (grade: A): "Safe Navigation in Dynamic Environments Using Control Barrier Functions and an RGB-D Camera". Supervised by Prof. Kostas Alexis.
• Quadrotor that avoids moving obstacles using RGB-D, scene flow, and CBFs — running at 20+ Hz on a Jetson Orin NX.

• Building a humanoid robot from the ground up — FPGA acceleration, motor controllers, whole-body control, and AI integration.
• Early-stage team — broad ownership across the full stack, from PCB-level hardware to high-level intelligence.
• Working at the intersection of embedded systems, real-time control, and machine learning.

• Specialized RAG pipeline over Norwegian legal sources.
• Combined retrieval, reasoning, and document generation into one product.
• Owned backend infrastructure and frontend interface end-to-end.

• Built and published Moving Dots on Google Play independently — no team, no guidance.
• Polished Unity/C# game with real downloads and ad revenue.

• Automatically regulates seed and fertilizer output based on predefined field zones — GPS positioning, zone mapping, and real-time control.
• Runs on an Arduino with RTK GPS and a personal RTK base station for centimeter-level positioning accuracy.
• Built independently, integrating embedded hardware, GPS, and custom control software.

• Connected speech-to-text, LLM reasoning, higher-level planning, and ROS motor control.
• Language model reasons about what to do before executing — not just keyword matching.
• Practical embodied AI, not a chatbot demo.

• Built backend, infrastructure, deployment, and product plumbing from scratch.
• Analyzed how companies perform on AI search engines like Perplexity.
• Applied to Y Combinator. Left before the company continued as Jarts.io.