Can you catch an AI lying by reading its private thoughts?
I'm a backend and fintech engineer with six years building reliability-critical systems — payment rails, cross-border settlement, agentic trading. Alongside that, I do independent research on AI safety: how large language models deceive, how well deception can be detected, and whether a model's private reasoning can be monitored to catch it. I released this as DeceptionBench, a live cross-play benchmark over six frontier models. Currently an MSc student in Applied IT at Ural Federal University.
Featured research
DeceptionBench — “Among LLMs”A monitor reading only a model's private reasoning catches the liar — across six frontier models, in a live cross-play benchmark.
ρ = 0.89deception ↔ detection
AUROC 1.00monitor accuracy
96%intent leaks in reasoning
Recent
- 2026Released DeceptionBench, a cross-play benchmark for LLM deception & monitorability, as a preprint on Zenodo.
- 2026Joined Figo as founding engineer — core backend for a cross-border neobank.
- 2025Began MSc in Applied IT at Ural Federal University.
Publications
- [1]
- [2]Discrete Event Simulation of a Municipal Meat Production System: Bottleneck Analysis and OptimizationSustainable Innovative Development: Design & Management · 2026 · RSCI (eLIBRARY.ru)
Writing
- Unlocking Financial Potential: FinSight AI's promise for empowered financedecentralised finance & ai
- I'll change the worldpersonal
Selected projects
Experience
- Founding Engineer — Figocross-border neobank · stablecoin rails, card issuing
- Founder / Lead Engineer — Jevan Studioproduct & engineering
- Founding Engineer — Autospendstablecoin infra · agentic trading system
- Senior Backend Engineer — BentoNigeria's leading payroll platform