Evidence is scattered
Company registries, screening results, ownership charts, adverse media and conflict checks typically live in separate tools, spreadsheets and emails. Reviewers rebuild the picture for every client.
Portfolio project · Interactive prototype v1.0
Client due diligence and risk intelligence in one explainable workflow.
RiskIntel AI is a client-acceptance platform for professional-services firms. It brings corporate research, AML, sanctions and PEP screening, beneficial-ownership analysis and conflict & independence checks into one case file, scores risk with a transparent rule set, and routes every decision through human approval with a full audit trail.
No sign-in, no install. Runs entirely in your browser on fictional data.
From the demo dataset
Indicative risk band · 15 cases
- Low7
- Medium4
- High2
- Critical0
- Insufficient information2
- Entities
- 21
- Individuals
- 21
- Ownership links
- 28
- Evidence items
- 40
Business problem
Client acceptance is high-stakes, fragmented and hard to evidence.
Judgements are inconsistent
Without a shared methodology, two analysts can reach different conclusions on the same facts, and a strong score in one area can quietly mask a blocking issue in another.
Decisions must be defensible
Firms need to show who reviewed what, which evidence and policy version supported the decision, and why a client was accepted, escalated or declined.
RiskIntel AI's answer: one case record shared by every module, deterministic and explainable scoring, non-compensating "gates" that stop a case regardless of its score, and immutable assessment snapshots tied to a versioned policy.
How it works
From intake to an auditable acceptance decision.
- IntakeRegister the client, engagement and related parties
- Corporate researchProfile, officers, shareholders, filings, litigation, media
- ScreeningSanctions, PEP and adverse-media candidates with explainable match scores
- OwnershipLook-through to beneficial owners, control indicators, gaps
- ConflictsRelationship matching against the firm's client and engagement register
- Risk assessmentWeighted score, gates, controls and residual view
- Human approvalRole-checked workflow; AI never decides
- ReportCompliance dossier built from the same validated record
Core capabilities
Ten working sections. Each card opens that module in the demo.
Executive dashboard
10 KPIs and 10 charts computed from case records, with six filters and click-to-filter charts.
Open →Corporate intelligence
Entity search and profile: officers, shareholders, financials, filings, litigation, media and evidence history.
Open →AML & screening
Sanctions/PEP candidates scored with Jaro-Winkler and token-sort matching; six human dispositions with role rules.
Open →Ownership intelligence
Path enumeration, look-through, voting vs. economic interest, cycle detection and an interactive ownership graph.
Open →Conflict & independence
Layered relationship matcher, restricted-register masking, threat mapping and review outcomes.
Open →Risk assessment
Eight weighted dimensions, coverage checks, blocking gates, controls, residual risk and immutable snapshots.
Open →Case management
Workflow states, tasks, prerequisites with reasoned overrides and role-checked transitions.
Open →Compliance reports
17-section acceptance dossier with an integrity guard; copy the text or print / save as PDF.
Open →Administration
Roles, versioned risk policy, thresholds, screening datasets, workflow mapping and audit log.
Open →AI assistant (simulated). The "AI assistant" panel shows how drafted summaries, explanations and executive summaries would appear, citing evidence records where the source data supports it. In this demo it is template-based: no language model or external API is called, and every output is labelled as requiring analyst review.
Analytical methodology
Explainable by design: every number can be traced.
Weighted risk model
Eight dimensions, weights from the demo's default policy (editable as a new policy version in Administration).
AML / sanctions / PEP20%
Client integrity / governance15%
Ownership transparency15%
Adverse media / legal / regulatory15%
Jurisdiction / geography10%
Industry / service / engagement10%
Conflict / independence10%
Firm capability / quality5%
Low 0–24.99Medium 25–49.99High 50–69.99Critical 70–100
Techniques in the engine
- Name matching: Jaro-Winkler similarity combined with token-sort comparison, with alert tiers (priority ≥ 85, review ≥ 65, low-priority ≥ 45).
- Ownership: deterministic path enumeration with multiplicative look-through, voting vs. economic interest, configurable UBO thresholds, cycle detection and gap accounting.
- Non-compensating gates: BLOCK / ESCALATE / HOLD rules (e.g. unresolved sanctions candidate, ownership opacity) override a favourable score.
- Coverage: missing evidence lowers confidence and can yield "Insufficient information" rather than a false Low.
- Governance: role-based permissions, immutable assessment snapshots tied to a policy version, and an audit log of every action.
Technical implementation
What this demo is built with, and what is planned.
This demo (implemented)
- Frontend
- React 18 single-page application, compiled to one static JavaScript bundle; hash-based routing across 10 sections
- State
- In-browser reducer over a shared case data model; persisted to this browser's local storage and resettable in Settings
- Engine
- Deterministic scoring, matching and ownership logic running client-side
- Charts
- Hand-built inline SVG components (no chart library), themed for light and dark mode
- Data
- Synthetic fixtures bundled with the app: 21 entities, 21 individuals, 26 officer records, 28 ownership observations, 10 screening scenarios, 9 adverse-media items, 10 conflict scenarios, 15 cases, 40 evidence records
- Hosting
- Cloudflare Pages, static Direct Upload; strict Content-Security-Policy that blocks all outbound network requests; self-hosted fonts
Production architecture (specified, not built here)
- UI
- Streamlit multipage application
- API
- FastAPI with versioned /api/v1 endpoints
- Database
- PostgreSQL with SQLAlchemy and Alembic migrations
- Agents
- LangGraph orchestration with schema validation and mandatory human review
- Analytics
- Pandas, NetworkX, Plotly
- Reports
- Server-side PDF (WeasyPrint / ReportLab) and Excel (openpyxl)
The project is specified across 15 design documents, from the Master Foundation Blueprint and Technical Implementation Specification to module specifications for corporate intelligence, AML/sanctions/PEP, ownership/UBO, conflicts and risk assessment, plus testing and deployment handbooks.
Suggested walkthrough
See the core of the product in about three minutes.
- Executive dashboardScan the 10 KPIs, then click a bar in "Case status distribution" to filter the whole page.
- AML & screeningPick CASE-002 Bluewater Trading FZE in the case selector and open a sanctions candidate to see how its match score is explained.
- Ownership graphFollow the look-through to beneficial owners; click nodes and edges, or switch to the table view.
- Risk assessmentSee the weighted score next to the blocking gates that override it, then compare inherent and residual risk.
- Compliance reportReview the 17-section dossier built from the same record; print or save it as PDF.
- AI assistantOpen "AI assistant" in the top bar and choose "Explain corporate ownership" (note the evidence citations), then "Prepare an executive summary". Simulated and template-based.
Tip: switch role in the top-right menu (e.g. Risk Manager, Administrator) to see permissions change. Settings → "Reset demonstration data" restores the original dataset.
Demonstration environment
Safe to explore. Nothing here is real.
All organisations, people, registry numbers and media items are fictional; web links point to reserved .example domains. Scores and thresholds are illustrative and are not a regulatory standard. Not affiliated with any firm.
- Demonstration environment
- Synthetic data only
- Read-only static deployment: no server, no database
- No production connectivity or external APIs
- No live customer information
- Your changes stay in your browser and can be reset
Project credit
RiskIntel AI
A portfolio project by Siddharth Naik, covering product specification, risk methodology, data model and the interactive prototype.