Problems, methods, and code you can run
Each case follows the same five blocks: the business question, the data, the method and why that one, the result, and the code. What differs is what can be shown. The first is fully public — repository, tests, example reports. Two describe doctoral research under peer review, so method is here and results are withheld until the journals rule. The last runs on synthetic data, because the engagement behind it is under NDA. Each page says which it is, at the top.
Multi-Agent Commodity Analyst
Four LLM agents decide; a deterministic econometric core in R computes. A guard stops the system from writing any number the models did not produce.
LLM Agents · MSGARCH · Python + R
Commodity Price Forecasting
Regime-switching volatility and time-varying higher moments applied to grain prices, benchmarked against the models a trading desk actually uses.
GAMLSS · Bayesian MSGARCH · Backtesting
Multi-Objective Portfolio Optimization
Evolutionary algorithms on the risk–return frontier for commodity portfolios, compared head to head rather than assumed.
NSGA-II · Differential Evolution · Pareto
Causal Modeling of NPS
What actually moves customer satisfaction, separated from what merely correlates with it — for a large corporation.
Causal inference · IPW · Synthetic data
Framework or Hand-Rolled?
The same multi-agent system twice — a hand-written loop and a LangGraph StateGraph. Both emit a byte-identical report; the difference is measured.
LangGraph · StateGraph · Parity testing