Karl Backan

Computer Engineering student at Mid Sweden University. Focused on ML systems, NLP research, and autonomous-AI tooling.

Currently finishing a Master's thesis on Swedish-NLP ticket classification.

View Projects

Live Status

00:00 AM

Sundsvall, Sweden

01 About Me

The Journey

I started writing code at age 10 in Roblox Studio. Self-taught from games into web, then full-stack, and now into machine learning at Mid Sweden University.

ii.

The Thesis

My Master's thesis compares three approaches to classifying Swedish IT support tickets: KB-BERT fine-tuning, HDBSCAN+UMAP clustering, and few-shot LLM prompting, with weak supervision via Snorkel.

iii.

The Build

Duktio — a multi-vertical SaaS for Swedish service businesses (Next.js, Supabase, Stripe, Fortnox, Swish).

Hello World!

02 Selected Work

UMAP scatter showing clusters of Swedish IT support tickets

Master's Thesis

Compares KB-BERT fine-tuning, HDBSCAN+UMAP clustering, and few-shot LLM prompting for Swedish IT-ticket classification. Investigates weak supervision via Snorkel and Swedish-English code-switching as a classification signal.

30 ECTS · Mid Sweden University · in progress

Duktio

Multi-vertical SaaS for Swedish service businesses (tradespeople, cleaners, dog daycare, vehicle detailing, tattoo studios). Next.js 15 + Supabase + Stripe + Fortnox accounting + Swish payments. Mobile packaging via Capacitor with PowerSync offline-first sync.

5 verticals · single codebase · ROT/RUT-avdrag, Fortnox, Swish

Norway split into game zones using vertex-adjacency analysis

Strategic Territory Game

Strategic territory game built on real GeoJSON country boundaries. Flask backend + Leaflet frontend, with graph-theoretic adjacency calculation and large-country zone-splitting for play balance.

Flask + Leaflet · GeoJSON · zone-splitting algorithm

Grad-CAM heatmaps showing what a CNN looks at when classifying cats vs dogs correctly

Neural Network Fundamentals

Multi-layer perceptron implemented from scratch in NumPy with analytical backprop (gradient-checked), reaching 1.0000 test accuracy on iris. Lab 2 in progress: CNN cats-vs-dogs in TensorFlow + Keras with Grad-CAM error analysis.

DT086A · Mid Sweden University · VT 2026

Histograms comparing four post-processing strategies on public mouse-movement research data

Sally

Streaming pipeline that rotates samples from three public mouse-movement research corpora (DFL + Bogazici + Chao Shen, 62 contributors). Studies how multi-source signal blending affects per-user identifiability via random-forest baselines.

3 public corpora · 62 contributors · binary stream format · rotation 40s ±15s

Pipeline: screen frame → YOLO26 detection → ONNX inference → policy → ADB action with EasyOCR side-branch

Visari

Headless Python computer-vision framework for Android automation. YOLO26 object detection with ONNX inference, EasyOCR for text extraction, and device control through ADB plus scrcpy.

Python · YOLO26 + ONNX · EasyOCR · ADB + scrcpy

Pipeline diagram: input observer → trajectory model → event dispatcher → post-processing with feedback signal

Java Personal Project

Long-running personal Java project exploring input simulation, motion modeling, and event-driven systems. Continuously maintained codebase — 4,786 commits as of 2026-05-04.

Java · 4,786 commits · personal project

Rengsjö Farm

Small Next.js + Tailwind site for Rengsjö Farm, a Swedish farm. Product gallery, contact form, SEO-tuned.

Next.js 15 · React 19 · TypeScript · Tailwind · Headless UI

Live site GitHub

03 Where I work

04 Let's Connect

Reach me here.

Email is the fastest path. GitHub for code, LinkedIn for the rest.

karl_backan@protonmail.com