Building Vetnio — AI clinical documentation for veterinarians. Stockholm.
I work end to end — product, backend, infrastructure and the model layer. Day to day that means TypeScript and Python, AWS, Stripe, database schema design, and LLM systems in production.
Before Vetnio: MSc in Machine Learning at KTH. My thesis trained sparse autoencoders on GPT-2 MLP activations to pull dense representations apart into sparse, interpretable features — sweeping latent width from 512 to 4096 against the L1 penalty, then reading out top-activating tokens per latent to see what each one had actually learned.
LLM systems in production. Structured generation, speech-to-text, and the harnesses that keep model output trustworthy enough to put in front of professionals.
Speech. Speaker diarization, ASR, speech and speaker recognition — the thread most of the public work below runs along.
Interpretability. Sparse autoencoders and feature decomposition. The research question I keep coming back to.
- Daily — TypeScript · Python · React · Node · PostgreSQL · AWS
- Modelling — PyTorch · NumPy · Hugging Face · Weights & Biases
- Also fluent in — Java · C++ · SQL · Docker ·
Day-to-day work at Vetnio is private. What's public is the machine learning and speech work from KTH — written from scratch, not wired together from libraries.
| Repo | What it is |
|---|---|
| Deep-Learning | Neural networks from scratch in NumPy. One-layer → k-layer with batch normalization, a char-level RNN trained with AdaGrad, and a VGG/ResNet CIFAR-10 project past 90%. Gradients hand-derived, no autograd. |
| Diarization---Project | Speaker diarization on VoxData. Benchmarked state-of-the-art embedding models against each other, then built a custom diarization model and compared clustering strategies. |
| Speaker-and-Speech | DT2119 Speech and Speaker Recognition. MFCC feature extraction, HMM alignment and concatenation, and phoneme recognition. |
| NLP-Course | Language engineering, implemented rather than imported: CKY parsing, transition-based dependency parsing, n-gram language models, NER, random indexing, word2vec/GloVe, char-level LMs and neural machine translation. |
| Computer-Vision-course | Image analysis and computer vision. Fourier analysis and Gaussian filtering, edge detection, and segmentation via k-means, mean-shift and graph cuts. |
| Text-summarizer | Abstractive summarization on the wikiHow corpus — T5 fine-tuning, evaluation and an interactive demo. |
| AI-course | Classical AI: search and heuristics, logic-based knowledge representation, HMMs and Bayesian inference, planning and decision theory. |
| TSP-Problem | Traveling salesman heuristics in C++ under a hard runtime budget on Kattis — greedy construction plus 2-opt local search. |