Personal site · Seattle · May 2026
Nikolaj Hindsbo.
AI engineer. I came up in mechanical engineering, drifted toward machine learning, and now work on both sides of that line.
- Currently
- Armada
- Before that
- CMU · UW
- Based in
- Seattle, WA
Background.
Engineering mechanics at UW–Madison, then a master's in AI at CMU, now building things at Armada in Seattle. The mechanical background keeps showing up — I still think about deployment constraints, physical limits, and what actually breaks in the real world.
On the ML side, I'm most interested in time-series forecasting, running models on the edge (quantization, efficient inference, squeezing intelligence into small packages), and multimodal agentic systems that generalize across tasks.
Outside of work: skiing, running (training for marathons), hiking, paddleboarding. Mostly in the PNW, which turns out to be a good place for all of those.
If anything here catches your eye, inbox is open.
Selected work.
SCOPE
A Blender-based simulation and benchmark for evaluating language-driven PTZ camera agents. Tested 19 SLM+VLM combinations across 536 tasks. Published at HRI '26.
- Python
- Blender
- Qwen3
- Moondream
- vLLM
- LLM-as-Judge
Fine-tuning LLaVA for Web Agents
Group project at CMU. We took LLaVA, fine-tuned it on VisualWebBench, and pushed the open-model score up.
- LLaVA
- LoRA
- PyTorch
- VisualWebBench
Waste Classification on a Raspberry Pi 5
Vision classifier I squeezed onto a Pi 5 so it could run offline. Real-world stuff ended up mattering more than test accuracy.
- ONNX
- Raspberry Pi 5
- OpenCV
- INT8 quantization
Refueling Satellite
Senior capstone at UW. Designed and analyzed a concept for refueling satellites on orbit.
- SolidWorks
- ANSYS
- MATLAB
Directional Buckling for In-Pipe Locomotion
Soft robotics work. Designed a compliant leg that buckles in a known direction so a small robot can walk through narrow pipes.
- ABAQUS
- Silicone elastomer
- 3D printing
Beach Cleaning Device
First senior capstone. A sand-sifter for picking microplastics out of beaches.
- CAD
- Mechanical prototyping
- Field testing
↳ Click any row for the longer version. Source on GitHub.
What I follow.
- Time series forecasting — making predictions that are actually useful at scale
- Edge inference — INT8 quantization, ONNX, running models on hardware with no cloud
- Multimodal agents — systems that see, reason, and act across modalities
- Generalization — models that transfer to new domains without falling apart
- Robotics — from soft actuators to language-driven camera control
- Physical constraints as design inputs, not obstacles
- Systems that close the sim-to-real gap
- Skiing — Cascades when possible, elsewhere when not
- Running — working toward marathons
- Hiking & paddleboarding — PNW has good options for both