Maria Prokofieva Associate Professor · AI-native finance and accounting · FCPA
AI for the human side of money.
I build AI that reads the human behaviour behind financial numbers, in how people manage, audit, invest and spend money, and I publish the evidence that it works.
Money, as people handle it
Four ways people hold money, and the behaviour underneath all of them.
-
Under their control
Management and boards
How executives and boards report, disclose and talk about the money they run.
-
Under contract
Audit and assurance
How auditors check it, and how AI agents doing that work can themselves be verified.
- Specialist verifiers for agent governance: pilot code (certified evaluation and policy-area specialists on τ²-bench) Software, 2026
- Specialist verifiers for agent governance: pilot results (judge and specialist scores and certification on τ²-bench) Dataset, 2026
- The CMA Agentic Platform: Autonomous Asset Verification and Algorithmic Auditor Governance Article, 2026
-
Their own
Investment and behavioural finance
How investors read companies and react to what they disclose.
-
Spending it
Marketing and agentic commerce
Agent-to-agent transactions break existing marketing metrics.
Current work Outputs will be listed here as they are released.
Underneath all four: human behaviour
Digital behaviour, purchasing and addiction are not a separate field. They are the layer under how people manage, audit, invest and spend.
- Machine learning of digital traces to detect risk for behavioural addictions online Nature Reviews Psychology, 2026
Proof, not promises
Models are released with open code, data and evaluation results, so quality and impact are tracked, not promised.
-
Specialist verifiers for agent governance: pilot code (certified evaluation and policy-area specialists on τ²-bench)
Code to measure how much of an AI agent's work can be certified safe to skip human review, testing small policy-area specialist verifiers against one generalist LLM judge on τ²-bench.
-
Specialist verifiers for agent governance: pilot results (judge and specialist scores and certification on τ²-bench)
Judge and specialist scores on all 4,448 τ²-bench trajectories, with certification summaries and a checksummed results register.
-
Managerial AI Optimism in US Technology Earnings Calls: An Educational Dataset
Measures of managerial AI optimism and six AI-claim dimensions from the earnings calls of 20 US technology companies, with a human-coded benchmark of 456 passages.
-
FinAgent: multi-agent equity research on a small NVIDIA Nemotron model
Agentic AI that runs on a CPU: LangGraph agents research public companies from SEC filings, a fact-checker traces every number to its source, and a two-hour hands-on lab comes with it.
-
Cohort-harmonise: dictionary-first harmonisation and claim auditing across youth-development cohorts
Aligns constructs across five longitudinal youth cohorts from variable-level metadata only, with no access to participant data.
-
esem: Exploratory Structural Equation Modeling
R package for exploratory structural equation modeling, with a step-by-step companion paper.
Advanced AI, taught without dumbing it down
University units, ACSPRI postgraduate courses and the agentic AI lab: advanced AI for any audience, specialist or not, with nothing simplified away.
Latest
All 71 outputs-
Gamers Wave 1–3: questionnaire and Fitbit dataset, rebuilt from the raw survey exports (de-identified, adults)
-
Specialist verifiers for agent governance: pilot code (certified evaluation and policy-area specialists on τ²-bench)
-
Specialist verifiers for agent governance: pilot results (judge and specialist scores and certification on τ²-bench)
-
Managerial AI Optimism in US Technology Earnings Calls: An Educational Dataset
-
FinAgent: multi-agent equity research on a small NVIDIA Nemotron model
-
Cohort-harmonise: dictionary-first harmonisation and claim auditing across youth-development cohorts