Providence brings transaction monitoring, screening, fraud and case management onto one auditable record, so every decision a bank makes can be explained, defended and replayed.
Each engine judges alone, evidence lives in inboxes, and the audit trail is rebuilt after the fact. Scroll to see what changes.
Alerts from every engine roll up to the customer. Every view, decision and version is stored once and can be replayed.
On sampled cases the analyst records a view in their own words before the score and copilot appear. You can measure whether suspicion was formed by a person or only signed off.
Ownership and payment networks, a hypothesis board and an evidence locker with fingerprints. Facts that cut the other way are required before QA.
The copilot drafts, citing evidence for every sentence. Uncited claims block sign-off until someone cites or removes them. Hold, tipping-off lock and monitoring start on filing.
See exactly what was known on any date, which rule, list and model versions were in force, and who decided. External auditors verify integrity with an independent tool.
Analysts, the MLRO, the board and the engineers who keep it running each get a home built around their day, all on the same record.
My day brings tasks, deadlines and changes on your cases into one list. Customer 360 shows everything the bank knows about a customer, in order.
The MLRO inbox orders reports, holds, overrides, exits and access requests by legal deadline, across every workspace, and your deputy takes over automatically when you are away.
Risk indicators against appetite, regulatory exposure, customer impact and the cost of compliance. Aggregated, with no customer or case details.
Error budgets, incident command and disaster recovery for a platform that screens payments around the clock. Engineers see system data, never case content.
Fewest platforms that do the job: one stream and batch engine, one relational store, one search engine. Every rule, list, model, FX rate and calendar is versioned, and every decision is stored as a snapshot.
A property without a measurement does not count. Each guarantee has an engineering control and a daily proof.
Control totals per feed reconciled every 15 minutes. A break is a Severity 1 incident.
Exactly-once processing with idempotency keys, proven by daily gap scans and chaos tests.
No clean decision in time: instant payments are rejected, others held. Nothing is released unscreened.
Tagged canary transactions for every typology must alert. A missed canary is Severity 1.
Inputs, versions and outcome stored as an object-locked snapshot, recomputed nightly.
The AI writes the query and shows it to you. The database computes the answer. Every number comes from your data, never from the model.
SELECT booking_date, merchant, amount_eur FROM transactions AS OF '2026-10-16' WHERE party_id = 'CUST-0088213' AND mcc IN ('5411','5499')
Every module reuses the same data spine, case model and decision record. None introduces a second case system.
Monitoring with a tuning lab, in-house screening with a safety net, real-time fraud and payment controls.
Networks, counterparties, correspondent banks, document intelligence and cross-case link suggestions.
Filing to every FIU, exit decisions with a contestation dossier, and a governed authority-request pipeline.
Rule scorecard, rule sandbox, model register, QA and training, internal lists and horizon scanning.
Rules-based routing, absence-aware SLA forecasts and an automatic MLRO deputy handover.
A daily home for analysts, Customer 360, a transaction explorer with red-flag tagging, and handovers that write themselves.
One inbox ordered by legal deadline, a report register, findings tracker, risk assessment and a sourced annual report.
Risk appetite on one page, cost and capacity scenarios, customer impact, regulatory exposure and the board pack.
Error budgets, incident command with the DORA clock, disaster recovery, access reviews and an external audit export.
Every rule records the provision it implements, so a change in law becomes a list of tasks.
Ongoing monitoring that can prove what it implements and that it still works.
Report without delay, hold for the right working days, never tip off.
No solely automated decision with legal effect; data only where it is needed.
Models rank and draft; people decide. Every prompt and output is logged.
Tested failure modes, third-party register and exit plans.
Real-time screening, cases and reports run in the bank’s own data centres. Elastic batch work runs in an EU cloud on pseudonymised data, and re-identification keys never leave the bank.
The prototype grows one slice at a time. Each slice works end to end on synthetic data, and has to pass the same checks (no completeness breaks, no missed canaries) before the next one starts.
Data contracts, synthetic data generator, decision record.
Triage, investigation and filing on one case record.
Name and sanctions screening with explainable scores.
Rules, canaries and real-time holds on synthetic payments.
Country configuration and reporting adapters.
A sequence of work for a personal research build, not a delivery plan.
Providence is a prototype I am building in my own time, to explore how AI-driven detection can stay explainable, contestable and accountable to the people who carry the reporting duty.
Providence informs my proposed doctoral research on AI-driven detection and the EU AML reporting obligation.
It is not a commercial product. There is no pricing, no customers and no partner institutions.
Explore the clickable prototype ↗