MuseCogni

Time-Series Foundation Models

Pretraining large transformer architectures on heterogeneous financial time-series data. Investigating whether scale alone can capture regime shifts, volatility clustering, and cross-asset dependencies without explicit feature engineering.

Neural Hawkes Processes for Market Microstructure

Multivariate point process models with neural kernel functions, applied to tick-level order book data. The goal: disentangle cause from correlation in high-frequency price formation.

Efficient Off-Policy Evaluation in Non-Stationary Environments

Standard OPE breaks when the data-generating process drifts. We propose a sliding-window importance sampling estimator with regret guarantees under bounded distribution shift.

tick-tock · High-performance tick data pipeline

Rust-core with Python bindings. Handles 100M+ ticks/day with sub-millisecond aggregation. DuckDB backend for OLAP queries.

regime-sight · Market regime detection toolkit

Unsupervised regime classification using statistical features. No look-ahead. Works on any OHLCV feed.

All systems operational

Training cluster: idle  |  Data pipeline: streaming  |  Inference API: 12ms p99