MSc student in Artificial Intelligence, University of Agder
CFO, SafeMediAI
I build AI that shows its work: Tsetlin Machines whose decisions are logical clauses you can read, guardrails whose safety properties are proved by solvers, and certificates anyone can re-check. Outside the lab I build reinforcement-learning agents, data systems and apps that run in the browser.
Research at CAIR, the Centre for Artificial Intelligence Research, under Ole-Christoffer Granmo, who introduced the Tsetlin Machine. Master’s research: glassbox-lm, steps toward a language model whose knowledge is a set of readable clauses.
Thesis: Automatic Removal of Sensitive Content in Images and Documents for Efficient Archiving, with Rami Amer and Yasin Hessnawi, for Mattilsynet.
Best bachelor thesis at the Faculty of Information Technology and Electrical Engineering.
Norwegian AI startup behind DeclassifAI.
Routes a layer through named concepts and emits a re-checkable certificate on every forward pass. On synthetic ground truth: δ ≈ 0.002 with no measurable accuracy cost; backdoor detection at AUC ≈ 0.90.
Readable content clauses, bounded-memory temporal monitors and a tamper-evident certificate ledger, on a CPU. F1 0.827 on deepset/prompt-injections at 107 µs per check, in 12.5 KiB.
Safety properties proved or refuted by SAT and SMT solvers, with bypasses repaired by construction. Retraining stalls at 8 provable bypasses; constructive repair reaches 0, and accuracy goes from 0.69 to 0.71.
A glass-box, safety-shielded agent for ARC-AGI-3 that completes two public levels; and a graph Tsetlin Machine for Ames mutagenicity whose minimal-edit recourse succeeds for 95.5% of positives.
AlphaZero-style Chinese Checkers agent with a 7.5M-parameter network and Gumbel MCTS, plus prototypes for coalition-aware search, decomposed values and a wreath-equivariant network. 324 tests pass on every push.
A desktop environment that runs offline in a phone’s browser; and a sleep forecaster built on Borbély’s two-process model that reproduces all 11 published findings it was checked against, in 18.7 KB.
473,189 weekly disease reports stored as a PostgreSQL star schema, MongoDB buckets and a Neo4j graph, with Kafka and ksqlDB streaming, FastAPI and React.
Indoor heat-risk forecasting with equity-aware triage; textbook-to-Anki on local models; and a desktop IDE that shows what a coding agent is doing.
All papers, with BibTeX: anwardebes.github.io/#papers
Python, PyTorch, C++, Rust, TypeScript, React, Next.js, Tauri, Three.js and WebGL, FastAPI, PostgreSQL, MongoDB, Neo4j, Kafka, Docker, Linux, Git and LaTeX.
Every project above has its code public on GitHub. The PDF is generated from this page.