Why AI integration is a research problem
Companies don't struggle to adopt AI because the models are weak. They struggle because the answers can't be trusted, the data can't leave, and the infrastructure costs too much. Each of these has engineering and management sides, but at its core is a fundamental question that no amount of tuning answers. That core is what we work on.
Four obstacles we work on
- Answers that can't be trustedLanguage models state records and numbers that are not in the data, and they count unreliably.
- Structure that goes unseenA model can't tell which parts of a large network play the same role, yet that is what matters for risk, redundancy and drug targets.
- Data that can't leaveBanks, pharma companies and governments can't send private records to a cloud AI service.
- CostLarge datasets usually mean database clusters and GPU servers that most teams don't have.
Deterministic computation, complemented by AI
Computation
Mathematics and algorithms do the parts that must be right. The same input gives the same output, every time, so results can be checked.
AI
AI handles language: understanding what people ask, planning the steps, and explaining results in their own terms.
Research at the boundary
The question we study is where the line between the two should sit, so that each does what it is good at and the whole system can be trusted.
Research areas
Graph symmetry
Finding, exactly, which parts of a network are interchangeable. It reveals redundancy and unique weak points, and it is also the basis of how we compress large graphs.
Lossless compression
Storing very large knowledge graphs in a single small file that can be read in place, so that questions can be answered without a database server.
Neuro-symbolic reasoning
Combining language models with symbolic computation: the model understands and explains, the computation does the reasoning, and the answer is checked against the data.
Quantum AI, checked
Certifying when a quantum machine-learning model can be replaced by an explicit classical formula, with no quantum hardware and no quantum circuit simulation, so that companies know what they actually need before they pay for it.
Tools we have released
Parts of our research are open source, so that others can check them and build on them.
Fast, exact symmetry of real-world graphs
Real graphs are full of twins: nodes with exactly the same neighbours. GraphCrunch finds them in linear time, contracts them, solves the much smaller graph, and lifts the answer back exactly. It returns which nodes and which links play the same role, and the lift is proved, not approximated.
View on GitHub →Does your quantum AI model need quantum computing at all?
Quantum machine-learning models are usually run on quantum hardware or on quantum circuit simulators. For a fixed, deployed quantum model, deqcert builds an explicit classical formula for its output, with a certified error bound. When the certificate holds, that formula can be used in place of both the quantum hardware and the circuit simulation, on ordinary computers. It makes no claim about quantum advantage in general, only about the model in front of you.
View on GitHub →The proof behind GraphCrunch: A Short Proof of Twin-Quotient Decomposition and Orbit Lifting (preprint).