14.0 Benchmarking: eflow DNI (Bayesm) Capabilities vs Kor.ai - ravkorsurv/kor-ai-core GitHub Wiki

πŸ“Š 14.0 Benchmarking: eflow + DNI (Bayesm) Capabilities vs Kor.ai

Summary

This page captures insights from a demo call with eflow, who have partnered with DNI (Bayesm) to implement AI-powered surveillance. These takeaways inform the design of Kor.ai’s Bayesian platform, specifically its risk inference, UI design, and product roadmap.

βœ… Key Takeaways from the eflow + DNI Demo

Area eflow/DNI Capability Relevance to Kor.ai Action for Kor.ai
1. AI Risk Scoring of Alerts Uses Bayesm engine to assign probabilistic scores to alerts instead of binary rules. Matches Kor.ai’s use of Bayesian Networks. Continue pgmpy/Agena scoring with explainability + audit trails.
2. Cause-Correlation Logic Cross-signal reasoning is built to infer causality (not mere co-occurrence). Supports Kor.ai’s node cluster approach (e.g. Trade, Comms, News). Clarify causal paths in DAG structure.
3. Cross-Product Surveillance Covers multiple asset classes in unified logic. Matches modular DAGs with reusable cluster nodes across products. Mock test case for multi-asset abuse (e.g. FX + Commodities).
4. Market News Contextualisation Price movements explained using public events/news to avoid false alerts. Aligns with NewsEventMatch and ExpectedPriceMovement logic. Implement β€œexplained move suppressor” node.
5. Book Risk Profile Awareness Alerts weighted based on a trader’s book risk and positional context. Currently underdeveloped in Kor.ai. Add TraderBookRiskProfile input node; cluster with HR/PnL data.
6. Analyst UX – Fewer Clicks UI designed to reduce number of steps to triage alerts. Matches Kor.ai’s goal of streamlined, one-glance views. Prioritize rich, single-screen alert summaries + audit trail sidebar.

🧠 Conceptual DAG Snippet (Textual Representation)

[CommsPatternCluster] ─┐
[TradePatternCluster] ─┼──> [CauseRiskNode] ───> [InsiderDealingAlert]
[NewsEventMatch]     β”€β”˜           β–²
                                   β”‚
[ExpectedPriceMovement] β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜
                        β”‚
[TraderBookRiskProfile] ─────────▢

🧠 Additional Node Concepts

Node Name Purpose
TraderBookRiskProfile Captures context: trader exposure, risk appetite, open positions.
ExplainedMoveSuppressor Lowers alert score if spike justified by known events.
CrossAssetAbuseCluster Flag for scenarios that span products (e.g., equity + credit).

πŸ–₯️ Analyst UX Design Considerations

Feature Goal
Alert Card View Show cause chain, score, typology, case age, and current status on one screen.
Case Detail Panel Include rationale, key drivers, audit log, analyst notes inline.
Keyboard Shortcuts & Filters Enable quick triage via filter-by-score, age, typology.
In-line Explanations Show why the score is high (e.g. "news did not justify price move + unusual comms").

πŸ“Œ Next Actions

  • Integrate TraderBookRiskProfile into model input library.
  • Simplify alert review flow in UI mockups β€” fewer screens, richer cards.
  • Draft test scenario: risk-aligned insider dealing (e.g., high exposure + timed trade).
  • Highlight this benchmark comparison in internal roadmap decks.