From anomaly to audit-ready SAR in seconds, not hours.
Deterministic triage tiers with grounded drafts and human review before anything ships.
uv run python scripts/pilot_bootstrap.py --smoke· zero cloud spend- Structuring pattern across 14 synthetic transfers
- Counterparty cluster flagged in synthetic ledger
- Narrative grounded to 3 cited statutory stubs
1.0
Citation Grounding
Deterministic stubs. Zero LLM hallucinations in regulatory citations.
177 ms
p50 Pack Assembly
P9 benchmark verified on 10-case evaluation with cryptographic digests.
80%
Preparation Savings
Simulated investigator prep reduction against standard 60-min manual baseline.
Zero PII
Beyond Customer VPC
Runs entirely on customer-hosted infrastructure. Zero external model training.
Financial Crime Typologies
Pre-trained IsolationForest feature extractors tuned on synthetic transaction patterns. Zero generative hallucinations.
Sub-Threshold Smurfing
Detects structured cash deposits intentionally placed just below the $10,000 BSA reporting threshold.
- 4 deposits inside 90 minutes
- Amounts kept between $9,200 and $9,800
- Beneficiary account shared across branches
Rapid Velocity Burst
Catches accounts moving funds at 8.4x their historical 30-day baseline across rapid multi-hop accounts.
$94,500 moved in 6m across 3 hops
- $94,500 transferred in 6 minutes
- Hop count: 3 intermediate pass-throughs
- Immediate outbound withdrawal trigger
Unseen Corridor Transit
Flags first-seen transactions routing through high-risk jurisdictions without business trading history.
Zero 12-month commercial history
- No prior volume on corridor
- FATF increased-monitoring transit
- Entity registered under 60 days
Immutable SHA-256 Audit Defensibility
Every anomaly evaluation produces a cryptographic hash verified by independent regulator exports.
The 4-Stage Compliance Pipeline
How FinSentinel turns raw transaction records into audit-ready regulatory evidence without generative hallucinations.
Durable Ingestion
PostgreSQL RLS · Transactional Outbox
Raw events indexed atomically by (tenant_id, source, event_id). Duplicates are rejected with zero side-effects.
Anomaly Scoring
Calibrated IsolationForest · Feature Store
Unsupervised anomaly model evaluates 14 velocity and structuring features against prior temporal baselines.
Grounded Retrieval
pgvector · Statutory Precedent Stubs
Semantic search matches anomaly patterns to authoritative statutory guidance stubs with strict abstention.
Immutable SAR Pack
LangGraph · Audit Ledger · Print HTML
Assembles chronology, citations, draft narrative, and two-step sign-off into an immutable, exportable package.
PostgreSQL RLS · Transactional Outbox
POST /transactions with original currency and ISO 8601 UTC timestamp
Atomic insert into tenant schema with strict PostgreSQL row-level security
Server-authenticated principal required. Never trusts arbitrary client tenant headers.
{
"tenant_id": "tenant_apex_prod",
"event_id": "evt_20261012_8841",
"amount": "USD 9,400.00",
"channel": "branch_kiosk",
"idempotency": "201_CREATED",
"outbox_status": "DISPATCHED"
}Every stage executes inside the customer security perimeter. No telemetry or PII leaves customer infrastructure.
Transaction Facts
domestic metro corridorUSD 9,400 aggregate
cash deposits at branch kiosks
Synthetic Party Alpha
Synthetic Party Beta
6 transactions in 24h · x3.1 velocity vs prior USD 1,900 per day
- sub-threshold count:6 deposits below reporting line
- time clustering:4 deposits inside 90 minutes
- round figure avoidance:amounts set just under round marks
- counterparty reuse:same beneficiary across all items
Precedent Guidance
SYN-PREC-101 similarity high, queued for manual review
SYN-PREC-114 similarity medium, closed with no action after review
Synthetic history only
Regulatory citations
STUBIllustrative cash activity record note [regulatory-stub-v1] Stub only, paraphrase, not legal text
STUBIllustrative review workflow note [regulatory-stub-v1] Stub only, triage guidance placeholder
SAR narrative
Draft hidden until Simulate SAR Draft runs
Audit digest
computing
{"id":"case-structuring-01","severity":"HIGH","score":0.68}
Claim first, then draft: synthetic flow only
Built for review, not black boxes
Every tier is a triage priority for analyst review, never a fraud probability. Synthetic demo data only.
| Capability | FinSentinel | Generic black-box LLM |
|---|---|---|
| Scoring | Deterministic IsolationForest tiers | Opaque score |
| Citations | Grounded stubs with abstention | Invented references |
| Sign-off | Two-step human confirm | Auto action |
| Audit | SHA-256 digest trail | No provenance |
| Data boundary | Never leaves customer estate | Trains vendor models |
| Deployment | Customer VPC plus air-gapped license | Vendor cloud only |
- Row-Level Security
- No External Training
- Customer VPC Deployment
- Append-Only Audit
GitHub Developer & Pilot Quickstart
Run the entire synthetic compliance stack locally. No cloud account required, $0.00 external spend, and zero customer data leaves your perimeter.
$uv run python scripts/pilot_bootstrap.py --non-interactive --smokeEvaluates 6 synthetic AML cases locally with deterministic scoring and zero network calls.
_
Evaluating for institutional compliance or need an air-gapped HMAC evaluation token?
Includes reproducible test fixtures and tamper-evident audit ledger sign-off tools.
What could a pilot save your team
Drag the slider to match monthly case volume. Estimates use synthetic pilot measurements only.
160
hours saved per month
Based on 48 min automated synthesis per case
24.6
analyst-days freed per month
Calculated at 6.5 productive investigative hours/day
Assumes 60-min manual baseline and 80 percent simulated saving. Partners supply own baseline for final estimates.