Marian Helcl Øverli.
Applied physics & mathematics from NTNU, Trondheim
// 5 years Acoustic Consultant, Automation, Developer
// AI engineer / Data Scientist / Developer
Building LLM tools, RAG systems, and ML pipelines that solve real problems.
Open to full-time AI engineering roles — based in Norway, open to remote.
MovCrafter
A browser interior designer. Rooms, furniture, 2D & 3D, GLB/PNG export.
Learn Storytelling
A six-level course in storytelling for LinkedIn & Instagram, backed by neuroscience.
NeuroUX
A 12-week neuroscience-backed UX/UI learning framework. 84 daily sessions, streaks & workspaces.
German Learner
Flashcards, grammar, quizzes & hover-to-translate stories across CEFR A1–C2.
Acoustic Anomaly
Detect anomalies in audio with spectrograms and a CNN. Industrial fault detection.
Meow Master
Train a procedurally-drawn cat through tricks across seven rooms with reward mechanics.
Wolf Behavior
Classify resting, foraging & traveling states from raw GPS tracking data.
18 projects →
RAG Chatbot
Retrieval-augmented assistant over your own documents.
Image Classifier
Train and deploy a CNN classifier in a single notebook.
Sentiment Analyzer
Detect tone in product reviews with a fine-tuned transformer.
Paper Summarizer
Turn arXiv preprints into one-paragraph TL;DRs.
Scene Narrator
Describe a video scene-by-scene with multimodal models.
Speech Translator
Live translate spoken audio between languages with Whisper.
Languages
- Python
- JavaScript
- C#
- SQL
- CSS
- Bash
ML Frameworks
- PyTorch
- TensorFlow
- HuggingFace
- scikit-learn
AI Engineering
- LLMs
- RAG
- Agents
- API Integration
- Claude Code
Tools & Platforms
- Git & GitHub
- Linux
- Jupyter
Data
- Pandas & NumPy
- Matplotlib
- PostgreSQL
- Vector Databases
How Stateless LLM Calls Collapse to One Answer
500+ independent API calls to Claude Sonnet 4. Each one returned a variation of the same 52-year-old male cardiologist.
Comparing SVG Output From Claude, GPT, and Gemini
Three frontier models, one six-word prompt. Claude drew unique characters, GPT drew uniform figures, Gemini wrote the most efficient code.
AI Security Threats in 2026
Prompt injection, deepfakes, model theft, data poisoning, and supply-chain attacks — what AI builders are facing this year.
Frontier Model Updates
Anthropic, OpenAI & Google ship new flagships almost monthly. The pace isn't slowing.
Open Weights Catch Up
Llama, Mistral & DeepSeek narrow the gap with closed labs on reasoning benchmarks.
Agents, Memory, Tools
Long-horizon agents with persistent memory dominate the latest arXiv preprints.
Connect
with me.
Please contact me for AI discussions, opportunities, or just to connect with me.