Most AI today is a black box: it gives you an answer and asks you to trust it. I work on the opposite. I build models whose decisions are made of readable logic, so a human can open the box, see exactly why it said what it said, and check it.
- 🧠 Master's student in Artificial Intelligence at the University of Agder (UiA), working with the Tsetlin Machine group at CAIR
- 🔐 Bachelor in computer engineering (cybersecurity and programming) from NTNU Gjøvik, awarded the Eurekaprisen 2025 for best bachelor thesis
- 🚀 CFO at SafeMediAI, the Norwegian startup behind DeclassifAI
- 🤝 Open to collaborating on projects that help people and create real value
|
Measured steps toward a logic-based language model built from Tsetlin Machine clauses. Three studies with pre-registered hypotheses, results reported as measured, including the negative ones.
|
Interpretability by construction. Routes a layer through a certified Legible Bottleneck and emits a runtime Faithfulness Certificate that bounds what the named concepts cannot explain. Paper plus reference implementation.
|
|
A CPU-only, sub-millisecond guardrail for agentic AI: Tsetlin-logic content guard, online temporal behaviour monitors and a tamper-evident certificate ledger anyone can verify.
|
Tsetlin Machines that remember. Echo-trace automata and bounded-LTL literals for trajectory-level verification of AI agents. Catches the slow-roll attack that per-step monitors miss.
|
|
The first generative Tsetlin Machine: discrete diffusion over Boolean vectors with readable clauses at every step. Sampling, anomaly scoring and imputation from one model.
|
A Graph Tsetlin Machine that inherits BERT's attention edges as its graph. BERT is then thrown away and the student learns pure Boolean clauses: 94.77% on R8, fully interpretable.
|
More Tsetlin Machine work (click to expand)
| Project | In one line |
|---|---|
| veritas | Formally verified, competence-aware Tsetlin guardrail for language models |
| reins | Certifying, minimal-intervention runtime shield for LLM agents |
| lucid | Glass-box, safety-shielded agent for the ARC-AGI-3 benchmark |
| Clause-Driven-LLM | Clause-based routing for LLM agents with a SAT-verifiable receipt per decision |
| clauseguard | An auditable, SAT-verified safety layer for any LLM |
| ConfTM | Conformal prediction for Tsetlin Machines |
| FedTM | Federated Tsetlin Machines for multi-site learning |
| RobTM | Tsetlin Machines with a certificate on every answer |
| HGTM · Multi-HGTM · RGTM | Hierarchical, multi-view and recurrent Graph Tsetlin Machines |
| graphtm-cbr | Graph-walking hierarchical TM with counterfactual Boolean recourse |
| tsetlin-zoo | A survey of Boolean clause learners |
| 🖥️ vsclaude | A full IDE with an animated pixel-art AI coding companion: 200 distinct agent behaviours |
| 📱 PocketOS | A whole desktop environment running in your phone browser, offline and installable |
| 📚 Mneme | Drop in a textbook PDF, get an Anki deck out. Runs locally, nothing leaves your machine |
| 🔥 PyroPredict | Wildfire spread forecasting that feeds evacuation planning |
| 😴 SleepArchitect | Goes beyond telling you how you slept |
If something here is useful to you, a ⭐ on the repo helps others find it.