Continuous transaction integrity for fintechs. An autonomous compliance agent that audits 100% of transactions against AML rules, investigates every alert, clears the false positives itself - with a written justification retained for audit - and drafts the regulatory reports for the ones that matter.
Status: validation MVP. This is a smoke-and-mirrors demo built to validate the product with compliance professionals. Everything runs client-side on synthetic, anonymized data. No production infrastructure exists yet.
Fintech compliance teams drown in alert fatigue. Legacy transaction-monitoring platforms flag thousands of transactions; over 90% are false positives, and a human analyst still has to investigate each one, then spend ~45 minutes drafting a Suspicious Transaction Report for the few that are real. Sample-based month-end audits cover a fraction of activity and still consume whole days.
Legacy monitoring stops at the raw alert. AuditLens investigates it:
- 100% continuous coverage - every transaction is audited against the rulebook, not a 5% month-end sample.
- The agent clears its own false positives - each clearance comes with a plain-English justification memo, retained as an audit trail (a payday routine is recognized as a payday routine, not a velocity alert).
- One-click regulatory reports - when a case is real, the agent has already drafted the NFIU-ready STR: subject, evidence, grounds for suspicion, statutory basis. The compliance officer reviews and signs; nothing is submitted without a human.
- Zero-PII ingestion - hashed customer and account identifiers only. Names, BVNs, and card data never leave the institution.
In the demo week: 1,224 transactions audited → 38 alerts → 34 auto-cleared → 4 cases for human review, with 4 STR drafts ready to sign and ≈23 analyst-hours saved.
Open dashboard.html in any browser - it's fully self-contained (no build,
no server, no dependencies; IBM Plex loads from Google Fonts when online and falls
back gracefully offline).
| Screen | What it shows |
|---|---|
| Overview | The week's verdict: cases needing review, alert-resolution funnel, coverage stats, alerts by rule. |
| Case queue | Open cases and auto-cleared alerts, sortable by risk, recency, or amount. |
| Reports | Draft STRs awaiting officer sign-off, plus the auto-filed CTR. |
| Rules & policies | The five detection rules with thresholds and per-rule outcomes. |
| Case file | Click any case: the agent's justification memo, transaction evidence with running totals, risk score, and the draft STR as a ready-to-print paper document. |
| Rule | Pattern |
|---|---|
VEL-01 Beneficiary velocity spike |
4+ transfers to first-seen beneficiaries within 30 minutes |
STR-02 Structuring below CTR threshold |
Repeated sub-₦1m transfers to one beneficiary exceeding ₦5m in 72h (never auto-cleared) |
CBM-03 Cross-border profile mismatch |
2+ coinciding signals: IP/KYC country mismatch, new device, weak name match, night-time initiation |
DOR-04 Dormant account reactivation |
180+ days dormant, then ₦1m+ outbound within 24h |
CTR-05 Currency transaction threshold |
Single transactions ≥ ₦5m (individual) / ₦10m (corporate); CTR auto-drafted |
| Path | What it is |
|---|---|
demo/dashboard.html |
The demo - open it directly in a browser. |
demo/dashboard_template.html |
Dashboard source with an __AUDIT_DATA__ placeholder - see build step below. |
data/generate_data.py |
Seeded generator for the synthetic dataset and agent findings. |
data/transactions.csv |
1,224 synthetic, anonymized Nigerian fintech transactions with planted violations. |
data/audit_findings.json |
The agent's output: alert statistics and six full case files. |
docs/agent_prompt.md |
Prompt to reproduce the audit live with an LLM, for "is the AI real?" moments. |
docs/DEMO_GUIDE.md |
Talking points, demo walkthrough order, and outreach playbook. |
plans/outreach-plan.md |
Validation step 2: target list, messages, interview script, tracking. |
screenshots/ |
Ready-to-send PNGs of every screen. |
Rebuild the dashboard after changing the data (from the repo root):
python data/generate_data.py
python -c "import io; d=io.open('data/audit_findings.json',encoding='utf-8').read(); t=io.open('demo/dashboard_template.html',encoding='utf-8').read(); io.open('demo/dashboard.html','w',encoding='utf-8').write(t.replace('__AUDIT_DATA__', d))"The demo cites the Nigerian framework: the Money Laundering (Prevention and Prohibition) Act 2022 (₦5m/₦10m currency-transaction thresholds, 24-hour STR rendition to the NFIU) and the CBN AML/CFT/CPF Regulations 2022. The same architecture generalizes to other regimes (FCA, FinCEN) by swapping the policy layer. Citations were written for a mockup - have a compliance professional verify exact sections before relying on them.
- All records are synthetic. No real customer, transaction, or institution data appears anywhere in this repository.
- "AuditLens" and "Demo Fintech Ltd" are placeholder names.
- This is a validation artifact, not a compliance product. Draft reports require human review; nothing here constitutes legal or regulatory advice.