AI gives answers.
We give reasons.
AI solutions that are transparent by design.
Funded by
Pilot partners include
Our Solution
Modern AI systems are large, complex black boxes. They show almost nothing about how they reach a prediction. This makes them impossible to use in highly regulated and critical cases.
To solve this problem, we developed a new method in the Academy of Finland project XAILOG. It learns AI models directly from the data, and those models are interpretable by design. Our solution is a cost-efficient and transparent alternative to black-box techniques such as Random Forests and XGBoost. It keeps competitive predictive performance.
Training Data
Historical labeled data from your domain
ExplainedAI
Learns transparent rules directly from data
IF condition OR condition THEN class A ELSE class B
Interpretable Rules
Human-readable IF/THEN classifier
Predictions
Apply rules to new, unseen data
Examples
IF Uniformity of cell size ≥ 3.5 OR Bare nuclei ≥ 3.5 AND 1.5 ≤ Uniformity of cell size < 3.5 OR Single epithelial cell size ≥ 2.5 AND Uniformity of cell shape ≥ 3.5 THEN Malignant ELSE Benign
Insights From the Data
This 3-rule classifier comes from the Wisconsin Breast Cancer dataset. It flags a tumor as malignant in three cases. The first case is a highly uniform cell size. The second case is a large bare nucleus together with a moderate uniformity of the cell size. The third case is a large epithelial cell together with an irregular cell shape. Each rule catches a different sign of malignancy.
Accuracy: our solution reaches 96.4%. Random Forest reaches 97% and XGBoost reaches 97%. Our solution stays fully interpretable.
IF Duration ≥ 15.5 months AND Not all credits at this bank paid back duly AND Unemployed OR Duration ≥ 15.5 months AND No checking account or < 200 DM OR No checking account or < 200 DM AND Not all credits at this bank paid back duly AND No guarantor THEN Bad Credit ELSE Good Credit
Transparent Decisions
This 3-rule classifier comes from the German Credit dataset. It flags bad credit in three cases. The first case is a long loan to an unemployed customer who has not paid back all credits at this bank duly. The second case is a long loan to a customer with no checking account, or with less than 200 DM on it. The third case is a customer with no checking account or less than 200 DM, and no guarantor. This customer has also not paid back all credits at this bank duly.
F1 score: our solution reaches 60.7%. Random Forest reaches 59.4% and XGBoost reaches 58.9%. Our solution catches more bad credit cases.
IF Holds no stock options AND Works overtime OR Job level is entry level AND Works overtime OR Job level is entry level AND Holds no stock options THEN Likely to leave ELSE Likely to stay
Rare Events, Caught
This 3-rule classifier comes from the IBM HR Analytics dataset. It flags an employee as likely to leave in three cases. The first case is overtime together with no stock options. The second case is an entry-level role together with overtime. The third case is an entry-level role together with no stock options. The second case is the strongest one. It covers 156 of the 1,470 employees, and 53% of them left. Company-wide only one in six left.
F1 score: our solution reaches 46.7%. Random Forest reaches 23.7% and XGBoost reaches 45.1%. Our solution catches two thirds of the employees who leave. XGBoost catches one third.
Black Box AI Compared With Our Approach
Black Box AI
- Predictions hard to explain or audit
- Complex and large
- Requires heavy computing infrastructure
- Compliance risk under EU AI Act / GDPR
Our Solution
IF Income < 40k AND Employment < 2y OR Debt-to-income ≥ 45% THEN Deny Loan ELSE Approve Loan
- Every prediction has an explanation
- Short and human-readable
- Cheap to deploy and maintain
- Regulatory-compliant by design
Already Have a Model?
You do not have to replace your AI system. Feed the predictions of your model into ours. You get human-readable rules that explain its decisions. This meets the explainability requirements of the EU AI Act and GDPR, and you keep your model. You see what your model learned, you catch hidden biases, and you build trust before you deploy.
Your Black Box
Your existing AI model
Labeled Predictions
Data with your model's predictions
ExplainedAI
Learns rules that mimic your model
IF Risk score ≥ 0.7 OR Age < 25 AND Income < 30k THEN High risk ELSE Low risk
Human-Readable Explanation
Human-readable rules that explain your model
Benefits
Rivals Black-Box Accuracy
Comes close to the best known accuracy on structured data.
Compact Results
Far smaller and more readable than decision trees.
Mathematical Guarantees
Backed by rigorous proofs on convergence and accuracy.
Runs on a Laptop
No GPU required. Fast training, instant inference.
Built for Real Tables
Seconds from a raw table to finished rules on one laptop CPU. No GPU, no cluster.
Training time on well-known datasets
Each bar is one dataset as published. Every row is in the training. The rules use the 20 most informative conditions. The largest dataset has 11,000,000 rows.
What You Can Do With It
Predictions Shareholders Trust
Loan approvals, insurance claims, medical triage — when an AI makes a decision that affects someone's life, regulations demand an explanation.
EU AI Act & GDPR compliant by designDiscoveries Anyone Understands
Point our solution at your data and it gives you patterns your domain experts immediately understand. No data science degree required.
Turn data into actionable knowledgeRuns Anywhere
Our models are plain if/then rules. No cloud, no GPU, no ML framework required. Deploy them on edge devices, embedded systems, or microcontrollers.
Predictions in any environmentOur Team
A world-class team from Tampere University, advancing transparent and understandable AI through our research in AI and logic.
Antti Kuusisto
- Led the Academy of Finland–funded XAILOG project
- 20+ years of research in logic and AI
- Publications in NeurIPS, AAAI, and JAIR
Tomi Janhunen
- 25+ years of research in logic and AI
- Chairman of the Finnish AI Society (2022–2023)
- Multiple Best Paper awards
- Publications in IJCAI, AAAI, JAIR, NeurIPS
Jussi Lemiläinen
- 25+ years as an executive and entrepreneur
- Product development & productization
- Finance and international sales
Reijo Jaakkola
- Publications in logic and explainable AI
- Ernst Lindelöf Award for Master's thesis
- Two-time Quantum AI hackathon winner
Veeti Ahvonen
- Logic and modern AI models (neural networks, transformers)
- Work featured at NeurIPS and AAAI
- 7+ years of industry collaboration
Selected Publications
Peer-reviewed research underpinning the technology, conducted within the Academy of Finland project XAILOG.
-
Expressive Power of Graph Transformers via Logic
-
Logical Characterizations of Recurrent Graph Neural Networks with Reals and Floats
-
Why This and Not That? A Logic-based Framework for Contrastive Explanations
-
Explainability via Short Formulas: the Case of Propositional Logic with Implementation
-
Interpretable Classifiers for Tabular Data via Feature Selection and Discretization
-
Short Boolean Formulas as Explanations in Practice
Become a Pilot Partner
We are looking for organizations that want to test our solution.
Free
No fees, no commitment. Business Finland funds us.
Tailored
We work directly with your team to adapt our solution to your domain, data, and requirements.
Easy to Integrate
Delivered as a library, plugin, or API. Fits into any existing platform or workflow. No infrastructure changes needed.