Regime detection in illiquid frontier markets.
We propose a Hidden Markov Model approach to detect regime shifts in markets where standard volatility-clustering signals are unreliable due to thin trading. Tested across NGX, NSE, and BVMW.
Research
Notes and findings from work in progress.
We publish selected research as we ship. Topics span quantitative finance, ML systems, and the practical engineering challenges that come with both.
Reward hacking patterns in LLM agent benchmarks.
Cataloging reward-hacking strategies observed across 200+ trajectory annotations on Terminal-Bench. We propose a taxonomy and recommend evaluation patterns that resist gaming.
Latency budgets for production RAG.
Lessons from architecting a RAG system serving 10M+ daily requests. How to reason about end-to-end latency, what to optimize first, and where the surprising bottlenecks usually are.