Demonstration systems designed and built end to end by Dmitry Krutous, an ETF and asset-management domain expert in Greater Boston. These are research and portfolio projects: they are not offered as an investment service and are not investment advice.
Six specialist agents communicating over a priority-based event bus above a signal pipeline: a watchdog tracking eight risk dimensions, a cost-aware rebalancer that trades only when the value ratio justifies it or drift exceeds five percent, a compliance agent that screens every proposed trade against mandate caps, limits, and prohibited lists before execution and maintains an immutable audit trail, plus researcher, communicator, and strategist agents. Underneath sits a prediction layer using regime-switching ensembles across Transformer and Mamba architectures with sentiment and margin models, and research into Kolmogorov-Arnold Networks, Neural Controlled Differential Equations for irregular time series, GATv2 graph attention, and conformal prediction for calibrated uncertainty. Delivery was planned as a phased production rollout: policy configuration, unit tests, shadow-mode burn-in, user acceptance testing, compliance sign-off, runbooks, and hypercare.
A voice-first investment profiling prototype exploring conversational onboarding, with a multimodal brokerage-statement upload and a thematic personalization layer. Built as a demonstration only and never operated as a live service.
evals.finance — runnable, rubric-graded evaluations measuring how well large language models perform real asset-management workflows, including public evaluations of post-trade operations. Source at github.com/DimaMerc/finance-llm-evals.
Contact: welt.management.solutions@gmail.com · LinkedIn