Independent NeuroAI Researcher · AI Systems Architect
Neural dynamics · Causal identification · Technical AI safety
I study when an adaptive system's behavior supports an inference about its mechanism—and when the evidence cannot distinguish competing explanations.
My work spans computational neuroscience, cognitive architectures, and reliable AI systems. The approach is consistent: state the assumptions, test interventions, search for counterexamples, and preserve the limits of each claim.
CWC · BN-Syn · GeoSync · neosynaptex · KORPUS
Five selected projects, with source-pinned routes to a mechanism, reproduction guide, adversarial probe or counterexample, and claim boundary. A source reference establishes provenance; it does not establish that an experiment was executed or independently replicated.
Inspect the evidence · Challenge a result
CWC extends Andrej Karpathy's nanochat. Its research, instrumentation, and evaluation layer is distinct from the attributed upstream work.
A bounded, deterministic gate compares declared candidate models with supplied observations. It can propose a discriminating intervention, reject the declared model class, or abstain. Selecting one surviving model does not grant causal authority.
Inference protocol · Executable kernel · Falsification traces
Probabilistic generation proposes. Deterministic contracts govern acceptance.
Mechanism → assumption → falsifier → evidence → admissible decision.
I build inspectable model and tool boundaries: typed interfaces, bounded authority, reproducible runs, and explicit failure conditions. The standard is not a plausible answer, but an outcome that can be tested against a declared contract.
Methods, tooling, and evidence limits
Controlled interventions, ablations, matched baselines, dynamical simulation, property-based and mutation testing, deterministic replay, and source-level provenance.
Python · NumPy / SciPy · FastAPI / Pydantic · pytest / Hypothesis · Docker · Linux · Git / GitLab CI
CWC's causal candidates remain conditional on identifying assumptions. BN-Syn's numerical properties do not establish biological fidelity. The neosynaptex γ ≈ 1 marker is not, by itself, evidence of criticality. KORPUS distinguishes engineering checks from release authorization. Independent replication and deployment require separate evidence.
Independent research from Ukraine. Open to focused collaborations in NeuroAI, cognitive architectures, technical AI safety, and evidence-bound agent engineering.
Email · Telegram · Engineering services
⊛ Vita est certamine.