Oleg Danilkiff
Systems engineer by day; independent research on developmental transitions in small learning systems after hours.
Seventeen years of building and running distributed systems in banking and telecom taught me one method, and I apply it on both sides of the day/evening boundary:
- evaluate against measured constraints before scaling;
- declare kill criteria up front;
- keep decision records;
- make verification executable — CI-enforced numeric parity, reproducible labs.
By day that method serves production architecture; in my own lab it serves research on how internal structure forms in reservoirs, neural-assembly models, and tiny transformers, with network topology as the controlled variable and full-trajectory instrumentation as the approach.
Current work
Trajectory Atlas — a one-year, pre-registered research program: which properties of emergence-like transitions are universal across substrates, and which are artifacts of gradient descent. Three fully instrumentable model organisms (echo-state reservoirs, assembly calculus, small transformers), one shared measurement contract, and the namesake public deliverable — complete developmental records of training and in-context runs, with scripts that regenerate every figure from raw data.
Publications & reproducible artifacts
- Инженерия платежей (Payments Engineering), 2026 A Russian-language technical monograph on card payments and A2A systems end to end: EMV, ISO 8583, 3-D Secure, tokenization and HSM-backed cryptography, acquiring/issuing, scheme economics, SBP and Mir, plus the operational disciplines around a live payment product (KYC/KYB, AML, fraud, reconciliation, disputes). DOI: 10.5281/zenodo.19884844 · source repository
- Acknowledged-write loss in Redis A reproducible lab measuring loss of acknowledged writes under partition failover and power loss across Redis, Valkey and KeyDB; the loss is architectural (async replication + optimistic acks), mitigations quantified. repository
- Explainable model serving in Go A cgo serving path for a Python-trained LightGBM model with native SHAP reason codes; numeric parity with the Python reference enforced in CI. repository
Education
M.Sc., Applied Computer Science (2008).