Basmu
Cognitive Augmented Discovery Agent
CADA revalidates transaction-monitoring alerts and triage cases—so compliance, investigation, and MLRO teams can see whether suspicion is supported, incomplete, or points elsewhere.
Human review required · Five synthetic cases · Access code required
The Problem
Transaction-monitoring alerts and internal referrals capture only part of the available evidence. The bottleneck is deciding what to investigate, what to monitor, what evidence is missing, and whether a human reporting decision is warranted.
Introducing CADA
CADA supports defensible financial-crime decisions—for monitoring analysts, investigators, MLROs, KYC/CDD teams, QA, FIUs, and strategic intelligence—not a single “expert investigator” persona alone.
Positioning
Most institutions already run transaction monitoring, screening, and case management. CADA connects to those signals and adds the missing decision layer: competing hypotheses, beyond-the-flag omissions, coverage honesty, and disposition recommendations for investigation, monitoring, evidence requests, or filing—always with a human in authority.
Cognitive Architecture
Each stage shows what works now, what is conditional on data or backends, and what remains roadmap—not investor abstractions.
Builds a dossier from filings, narratives, subjects, and transaction maps when data materializes.
Reports what was found, failed, not queried, or incomplete—so gaps stay visible.
Deterministic typologies and subject graphs when the graph is available.
Competing illicit explanations plus a mandatory lawful / benign alternative.
Stress-tests the top story for gaps, defensibility, and provisional signal strength.
Recommends a reporting-class disposition with rationale and human review.
Adjudicated cases evaluate CADA; live scoring changes only under human-approved governance.
CADA sandbox
The protected sandbox contains five synthetic demonstration cases. Reviewers use an access code and can switch between the decision-date record and the full file. Each case has one active workspace tab at a time.
Read the customer profile, the activity under review, competing explanations, and the six investigation questions: who, what, when, where, why, and how.
See what supports or weakens a concern, inspect payment relationships, and trace the evidence linked to graph findings.
Compare the recommendation with the available evidence, record an analyst decision locally, and review a draft reason for suspicion. CADA does not file a report or determine guilt.
Case narratives and graph links are being reviewed for evidential accuracy before wider bank-facing use. Synthetic demonstration labels do not identify a real person or institution.
Open the protected sandboxRoles
Challenge: Alert floods and defensive STRs leave teams unsure whether suspicion is supported.
CADA: Argues the flag up or down; routes to investigate, monitor, request documents, or file—with coverage honesty.
Request DemoChallenge: Scarce investigation capacity cannot deep-review every lead.
CADA: Evidence-weighted dispositions and residual-risk documentation so MLROs can allocate attention defensibly.
Request DemoChallenge: Broad fishing for SOW/SOF wastes customer and analyst time.
CADA: Asks for the specific evidence that would confirm or refute a live hypothesis—tipping-off aware.
Request DemoChallenge: Incomplete packages and selection-biased predicates slow feedback.
CADA: Re-analyzes a triage case with competing predicates, omissions, and a feedback-ready structure while keeping the decision with the human reviewer.
Request DemoChallenge: Checkbox exams miss whether monitoring and filing are effective.
CADA: Can aggregate beyond-the-flag omissions and hypothesis themes to inform scenario tuning and thematic risk assessment—under institutional governance. Industry-wide scoring remains roadmap.
Request DemoChallenge: Opaque AI is hard to challenge, version, or promote safely.
CADA: Version stamps, blind-eval doctrine, shadow lessons, and human approval gates before live scoring changes.
Request DemoTechnology
CADA combines relationship intelligence with narrative reasoning so teams get findings they can defend—not opaque scores.
Subject, account, and transaction views when the graph materializes; multi-hop tracing where supported.
Structured hypotheses with a mandatory lawful / benign alternative.
Counterfactual checks on competing predicate stories when enabled.
Adjudicated cases evaluate CADA; live scoring changes only under human-approved governance.
Evidence chains, provisional signals, and audit-ready rationale.
Integrates with case systems via secure APIs and batch pipelines.
Explainability & Trust
Human-in-the-loop by design. Provisional calibration status shown per case. Nigeria legal pack validated; additional packs in validation.
Transaction flow showing why the pattern is treated as suspicious.
Each finding linked to subjects, accounts, and events.
NG-first pack citations to typologies and AML/CFT standards; other jurisdictions as packs validate.
Signals labeled provisional or uncalibrated when production probabilities are not available; abstention when coverage collapses.
Similar investigations surfaced only when institutional memory is queried and returns results.
Challenging and lawful hypotheses retained for transparency.
Learning & Evolution
Live scoring changes only under human-approved governance. Lessons stay shadow-only until promoted.
Baseline coverage on known pattern families
First novel pattern surfaced for review
Expanded pattern library from confirmed finds
Faster detection on recurring structures
Governed promotion into live strategies with human oversight
Security & Compliance
Full control; air-gapped option; no forced cloud egress.
AES-256 at rest · TLS 1.3 in transit.
Role-based, audit-logged authorization.
GDPR · SOC 2 · ISO 27001 readiness path.
Complete lineage of every decision.
Version control, fairness audits, rollback.
Deployment
Air-gapped capable. Custom integration with SIEM, case management, and core banking. Dedicated support and training.
Scalable infrastructure with managed options and multi-tenancy isolation per institution.
Secure access for real-time and batch pipelines with SaaS-style consumption where policy allows.
Proof Points
Surface what was found, failed, not queried, or incomplete—so gaps stay visible.
Prefer defensible leads and provisional signals over opaque risk scores. Validation pass bars are internal targets—not production precision claims.
Compress manual synthesis into guided review—without claiming minutes-to-hours certainty.
Targets above are design goals and early internal aims. Customer-approved results will replace them when available. No third-party endorsements are claimed. Sandbox cases are synthetic and non-live.
Roadmap
Team
Basmu is building CADA to make financial crime unprofitable through cognitive AI—guided by AML practice and investigator workflows.
Founder
Building Basmu and CADA — the Cognitive Augmented Discovery Agent for financial crime investigation. Focused on explainable discovery, investigator trust, and sovereign deployment.
LinkedIn profileNext Step
Request a demo tailored to banks, fintech, law enforcement, or regulators—or connect on LinkedIn.
Prefer email? amukhtar@basmu.xyz