Research Apps Websites About Contact View CV
← Back to portfolio

Publications & studies

Research

Studies where the data is messy and every answer is checked against something real: open payment data against public records, and unknown instrument data against physics.

01 2 studies · 2026
Bachelor's thesis · Published in DiVA

Automatiserad riskanalys av kommunala leverantörsfakturor

Nataniel Ataseven · Mid Sweden University · 15 ECTS

A seven-module Python system, four years of open payment data and a document-based investigation. Semantic screening led to a case with 13 documented control deficiencies — and a finding about the limits of external scrutiny.

5.4M transactions 107.6B SEK 13 control deficiencies
PythonPandasBenford's LawDesign Science
Explore the study →Publication in DiVA ↗
Open-source software · Pre-registered benchmark

Rosetta: naming unknown columns in battery‑lab data without a language model

Nataniel Ataseven · Independent project · 2026

When a lab instrument adds a new data column, open-source parsers stop reading the file. Rosetta reads it anyway and names each unknown column from physics, with the evidence. Tested on five frozen hold-outs against Claude Opus 5.5, it gave half as many wrong names and abstained where the data could not support a name.

373 real columns 5 wrong names vs 10 No language model
PythonNumPyBinary formatsBenchmark design
Explore the project →Source on GitHub ↗