Where ideas become
working intelligence.

FIBO-based ontology layer for lending and deposit — ask natural language questions over a knowledge graph backed by Jena Fuseki, Snowflake, and S3 Vectors. Powered by two custom fine-tuned SLMs (97%+ intent accuracy) deployed via Bedrock Custom Model Import.

Policy harmonisation engine for financial regulation. Ingests regulatory text, extracts obligations, maps controls, and tracks traceability from Regulation → Policy → Obligations → Controls. Built for NatWest / CNI onboarding transformation.

Neo4j knowledge graph wired end-to-end into the MidasLend engine. Explore relationship networks between entities, counterparties, and transactions. Natural-language queries over graph topology via Claude on Bedrock.

A pool of specialized agents that run the process around AI-assisted delivery — sprint planning, standups, retros, backlog refinement, impediment escalation, and stakeholder reporting — with Monte Carlo release forecasting and a live code-delivery view (PRs, CI pass rate, deployments). The governance layer that sits above coding agents like Kiro, not a replacement for them.

One engine, three surfaces: RM Desktop, Customer, Credit Ops. Resolves who the client is, what they already hold, what comparable clients hold, and what the gap is worth in pounds. All five layers built — L1 Data, L2 KYC.AI, L3 Ambient Capture, L4 Sales Surface, L5 Revenue Intelligence.

2.4M-parameter encoder model for business lending decisions. Treats every piece of evidence as a timestamped event stream — scores PD, debt service capacity, and early warning in under a second. Governed by the Aegis policy engine with signed, hash-chained Decision Records on DynamoDB.

Fine-tuned Llama-3.2-3B model for real-time financial communications surveillance. Classifies MAR/FCA violations, scores risk 0–100, and streams structured alerts — powered by PRAGMA tokenisation pipeline and Bedrock Custom Model Import.