WtrDB for Banking
Knowledge graph infrastructure for financial institutions — grounded in the lessons of JPMorgan Chase's JEL entity linking system.
Built from the same conviction as JPMorgan's JEL.
JPMorgan Chase published one of the most honest papers in enterprise knowledge graph literature when they described JEL — their neural entity linking system for financial news. The paper admitted that existing academic tooling, trained on Wikipedia and tuned for clean data, simply broke when applied to the messy, high-stakes reality of enterprise finance. They built their own entity embeddings, their own character features, and their own Wide & Deep architecture because no off-the-shelf solution could handle the ambiguity of company names in financial news. WtrDB was built from the same conviction: that banks need knowledge infrastructure that earns trust through governance, mathematical rigor, and auditability.
What banks actually need from knowledge graphs.
Entity Resolution at Scale
JEL opens with a deceptively simple example: the name "Lumier" appears in two financial news articles — one refers to a software company, the other to an LED manufacturer. Both are financially relevant, but to different stakeholders. If entity linking fails, the wrong credit officer gets the alert. JPMC reported maintaining several million entities and several million links. The scale demands infrastructure that handles ambiguity, evolving relationships, and continuous ingestion from heterogeneous sources. WtrDB addresses this through its 13-type entity taxonomy, dual-track extraction (static triples + evolutionary events), and three-filter governance pipeline.
The Wikipedia Problem
All existing entity linking systems are trained on Wikipedia. Wikipedia does not cover every entity of financial interest — the startup "Lumier" that raised investment from prominent VCs had no Wikipedia page, but it absolutely mattered to JPMC's risk and investment teams. WtrDB does not depend on any external knowledge base. It builds its own knowledge graphs from the institution's own documents. Every entity embedding, every relationship extraction, every governance decision is grounded in the institution's own data.
Context Poverty
JEL showed that context similarity methods performed poorly because descriptions in internal KGs use different language than Reuters or Bloomberg. Comparing context words results in low accuracy. WtrDB's hybrid retrieval combines vector similarity over chunk embeddings (captures semantic proximity regardless of wording) with typed graph traversal. When one signal is weak, the other compensates. Graceful degradation means if the graph layer is unavailable, the system falls back to vector search rather than failing.
Explore and rotate your WtrDB graph in 3D
Drag to rotate. Scroll to zoom. Right-click and drag to pan. Click a node to inspect its relationships.
Eight core use cases.
Supply Chain Risk Intelligence
JPMC's Acma Retail scenario — a retailer files for bankruptcy, stress propagates through the supply chain, different clients face different risk levels depending on their order in the chain. The system must accurately link "Acma" to the right entity and traverse the supply chain graph.
Financial News Analytics and Alert Generation
JPMC needed to distinguish "Acma Global Retail Inc" from "Acma Furniture, LLC" and "Acma Enterprise System" so alerts reach the right stakeholders. Before JEL, Tri-Gram Cosine Similarity had practical limitations and could not distinguish entities sharing similar names.
Regulatory Compliance and Audit Readiness
Banks operate under Basel III/IV, CMMC, SOC 2, GDPR/DPDPA, and jurisdiction-specific regulations that change constantly. Every regulatory examination demands evidence. Every audit requires a trail.
Anti-Money Laundering (AML) and Sanctions Screening
AML requires understanding complex ownership, control, and transaction flow webs. Sanctions screening demands matching against government lists (OFAC, EU, UN) while handling name transliterations, aliases, shell company structures, and beneficial ownership chains across jurisdictions.
Credit Risk Assessment
Credit risk teams synthesize financial statements, credit bureau reports, industry analyses, news coverage, internal relationship history, and macroeconomic indicators — arriving in different formats, using different terminology, sometimes contradicting each other.
Investment Research and Advisory
Research analysts synthesize earnings calls, SEC filings, industry reports, competitor analyses, and proprietary research. The challenge is connecting insights across sources, maintaining consistency across research teams, and ensuring recommendations are grounded in auditable evidence.
Fraud Detection and Investigation
Fraud investigation is a graph problem. Fraudulent networks involve rings of accounts, shared addresses, common beneficiaries, coordinated transaction timing, and layered corporate structures designed to obscure true parties.
Know Your Customer (KYC) and Client Onboarding
KYC requires identity verification, beneficial ownership understanding, risk profiling, and maintaining records as customer circumstances change — across passports, utility bills, corporate registrations, financial statements, sanctions lists, and PEP databases.
WtrDB as the knowledge graph operating environment JEL assumed.
JEL achieved 99.91% F1 accuracy by combining Wide (character pattern) and Deep (semantic embedding) learning. But JEL solved one specific problem: linking company name mentions in news to entities in JPMC's knowledge graph. The broader challenge remains — how do you build, govern, and continuously update the knowledge graph that entity linking operates against?
WtrDB handles everything before and after the entity linking step:
Before entity linking
- Ingest the documents containing mentions.
- Chunk, embed, queue for extraction.
- Pull entities and relationships through governed incremental processing.
- Adjudicate every candidate fact through evidence, logical, and evolutionary-intent filters.
- Write the validated graph to FalkorDB with full provenance.
During entity linking
- Hybrid retrieval combines vector similarity (semantic matching, analogous to JEL's Deep component) with graph traversal (structural context for disambiguation).
- Natural language to Cypher translation for analyst queries.
After entity linking
- Measure graph quality via sheaf theory.
- Detect conflict cycles.
- Enable federation across knowledge bases.
- Publish to external agents via Brain Endpoints.
- Maintain the governance trail regulators require.
JEL's identified limitations — directly addressed
“Lack of context information”
→ WtrDB builds context organically through multi-document ingestion. Every document adds context to the entity's node over time.
“Different wording styles”
→ text-embedding-3-small generates semantic embeddings that capture meaning regardless of surface language. Hybrid retrieval compensates when wording diverges.
“Blocking complexity”
→ Becomes a graph traversal problem: plausible entity matches found via both semantic similarity and structural relationships.
Built for regulated environments.
Multi-Tenant Organization Structure
Organizations, teams, members — per-tenant configuration and quotas. A global bank can deploy separate knowledge bases per business unit while federating for cross-unit intelligence.
Role-Based Access Control
System roles + custom org-scoped roles + team-scoped assignments. Junior analysts get viewer access to the full graph, editor access only to their team's KB. Compliance officers get read access across all KBs. Full audit trail in rbac_audit_events.
Data Architecture
PostgreSQL for relational metadata, audit trails, billing, and RBAC (transactional integrity for regulators). FalkorDB for graph data (query performance for intelligence workflows). Each store scales and recovers independently.
Webhook Integration
HMAC-SHA256 signed webhook subscriptions with exponential-backoff retry and delivery logging. Integrates with enterprise event buses. New entity extracted, relationship deprecated, or compliance artifact generated → downstream systems receive signed, verifiable notifications.
Four phases to full deployment.
Foundation
Months 1–3- Deploy WtrDB core infrastructure (PostgreSQL, FalkorDB, Redis).
- Configure RBAC, MFA, compliance framework auto-setup.
- Establish first KB for a contained use case (e.g. supply chain risk for one business unit).
- Ingest initial corpus and validate extraction quality against known entity relationships.
Intelligence Operations
Months 4–6- Scale ingestion to continuous news feeds, regulatory filings, internal documents.
- Deploy sheaf engine for quality measurement.
- Configure Brain Endpoints for first agent integration (e.g. credit risk alerting).
- Establish federation between initial KB and second business unit's graph.
Enterprise Scale
Months 7–12- Extend federation across business units.
- Deploy The Void for investigation teams.
- Integrate OPA policies for regulatory compliance enforcement.
- Establish scheduled compliance reporting.
- Build Brain Endpoint integrations for multiple agent consumers.
Platform Maturity
Year 2+- Optimize sheaf engine parameters for institution-specific consistency requirements.
- Deploy cross-jurisdictional federation for global operations.
- Establish the knowledge graph as the institution's canonical intelligence layer, with Brain Endpoints as the standard interface for AI agent development.
Against every alternative.
Against Traditional RAG Solutions
Most RAG products retrieve document chunks and pass them to an LLM. No governance, no entity resolution, no contradiction detection, no mathematical quality measurement. WtrDB provides all of these. Three-filter governance ensures facts are evidence-backed. Sheaf engine measures and maintains consistency. Provenance trail ensures every answer traces to its source.
Against Generic Knowledge Graph Platforms
Neo4j, TigerGraph, Amazon Neptune provide graph storage and query capabilities but do not include document ingestion, LLM-based extraction, governance adjudication, mathematical quality measurement, or agent-facing publishing. These are graph databases. WtrDB is a knowledge graph operating environment. The difference is between buying lumber and buying a house.
Against Point Solutions
Entity linking solutions, compliance tools, and fraud detection platforms each solve one piece. WtrDB provides the unified knowledge infrastructure that connects all of them. The same governed graph powers entity linking, compliance evidence collection, fraud investigation, credit risk assessment, and investment research. One infrastructure. One governance model. One audit trail.
Ready to bring this to your institution?
The conversation would be focused, specific to your context, and no-obligation.
This page references JPMorgan Chase's JEL research paper (Ding et al., AAAI 2021) for educational and comparative purposes. WtrDB is a proprietary product of Sftwtrs.AI (Nilesh AI Systems Pvt Ltd).