Technical architecture

Built for proof, not for demo.

The Settlement Control Plane is 20,277 lines of Python — no third-party ORM, no framework coupling, no black-box regulatory logic. Every detection module is tested against real adversarial datasets. The regulatory citation is the test specification.

120

pytest tests passing

15/15

adversarial datasets

2.26s

for 100K transactions

20,277

lines of Python

Technology stack

Purpose-selected. No framework debt.

Each layer was chosen for the constraint it solves — not to match a job posting or a VC's preferred stack.

Core language

Python 3.11+

Chosen for the financial-math ecosystem and for the regulatory citation test pattern — the spec is the test, not a separate document.

pandas · numpy · pydantic
decimal · datetime

API surface

FastAPI

Schema-first REST — the OpenAPI spec is generated from the same Pydantic models that enforce regulatory field constraints at runtime.

uvicorn · httpx
openapi-schema-pydantic

Data layer

SQLite + Pandas

Single-file persistence for the settlement ledger. Pandas for batch reconciliation windows. No ORM — SQL is the specification language.

sqlite3 · pandas
sqlalchemy (schema-only)

Validation

Pydantic v2

Every settlement message and compliance event is validated at the boundary. Regulatory field constraints are Pydantic validators — they cannot be bypassed by caller code.

pydantic · annotated-types
pendulum (timezone-aware)

Test framework

pytest + hypothesis

120 tests; 15 adversarial datasets. Hypothesis generates edge-case transaction sets that catch off-by-one errors in SLA window calculations.

pytest · hypothesis
pytest-cov · pytest-asyncio

Deployment

Docker + Compose

Single-container SaaS or multi-container on-premises. The connector adapter layer is the only deployment-variable code — the detection core is identical in both.

docker · docker-compose
gunicorn · nginx

Validation proof

The test suite is the proof of regulatory compliance.

Tests are named by their regulatory citation. The test passes when the detection matches the statutory obligation — not when it satisfies a product spec written by a product manager.

120 / 120

pytest tests passing · 0 failures · 0 skips

The full detection suite runs on every commit. Tests cover the happy path, the SLA boundary (T − 1 day), the SLA breach (T + 1 day), and the adversarial case (malformed input designed to produce a false negative).

15 / 15

adversarial datasets · all passed

Each dataset targets a specific regulatory edge case: an OFAC match with a one-character name variation, a batch with a sub-penny reconciliation gap, a Reg E dispute that arrives at 23:59 on day 10. All 15 are caught.

2.26s

for 100,000 transactions — full detection pipeline

Full 6-stage INGEST → MATCH → DETECT → CLASSIFY → GOVERN → REPORT pipeline at 100K-transaction scale. Scales linearly to the connector's ingest rate — the bottleneck is the network, not the detection engine.

20,277

lines of Python · 0 lines of framework config

No YAML configuration that hides logic. No ORMs that hide queries. No framework DSLs that hide the regulatory intent. The regulatory obligation is the code — readable, auditable, and citeable.

Detection pattern

The citation is the test specification.

# BSA 31 CFR §1020.320 — SAR filing within 30 calendar days of detection.
# Test: suspicious pattern detected on day 0; SAR must be filed by day 30.
# Adversarial: detection event timestamp is 23:59:59 UTC (boundary case).

def test_sar_30d_boundary_exact():
    event = SuspiciousActivityEvent(
        detected_at=datetime(2026, 1, 1, 23, 59, 59, tzinfo=timezone.utc),
        pattern="velocity_layering",
        amount_usd=Decimal("12500.00"),
        counterparties=["ACCT-8821", "ACCT-3094"],
    )
    deadline = sar_filing_deadline(event)  # must return 2026-01-31
    assert deadline == date(2026, 1, 31)

# OFAC 31 CFR Part 560 — same-day freeze on SDN match.
def test_ofac_sdnmatch_same_day_freeze():
    txn = Transaction(counterparty="AL BARAKAAT INT FOUNDATION", amount=Decimal("50000"))
    result = ofac_screen(txn, sdnList=SDN_LIST)
    assert result.match == MatchLevel.CONFIRMED
    assert result.freeze_triggered       # same-day, §560 — not deferred
    assert result.notice_package_ready   # OFAC notification assembled

Design invariants

What cannot be changed without breaking the product.

Invariant 1

Regulatory citations are structural, not configurable

Invariant 2

Detection output is citation-anchored — every finding references its statutory source

Invariant 3

SLA clocks start at event receipt, not at queue processing time

Invariant 4

No participant's data is visible to another participant through any API surface

Invariant 5

The method is never exposed — output is the right to run, not the source

Invariant 6

Connector adapters are the only deployment-variable code — the detection core is identical on every deployment