Origin Story · Know Your Community (KYComm)
Every piece of infrastructure starts as a question. For KYComm, the question came from an academic paper: who pays the price for financial-crime algorithms? The answer was the very people financial inclusion exists to serve. That finding set sixteen months of work in motion. Scroll the story the way it happened: one milestone at a time.
The paper that named the problem
Katie Grillaert authors Risks to Financial Inclusion by Anti-Money Laundering and Financial Counterterrorism Algorithms, examining how detection algorithms exclude the very populations they protect, and identifying federated learning as the mitigation.
Published by IEEE (ETHICS 2025), it remains the peer-reviewed foundation of this public good initiative.
Where should the project sit?
Numerous meetings with the Presidential Committee for Economic and Financial Inclusion (PreCEFI) and with Nigeria’s regulators and agencies, the Central Bank of Nigeria, the National Insurance Commission and the Nigeria Financial Intelligence Unit, to establish where the KYComm infrastructure should be domiciled and how it should be governed.
MOU with the Responsible AI Governance Institute
Goose FL signs a memorandum of understanding with the Responsible AI Governance Institute (RAI-GI), an Abuja-based public-interest nonprofit, as local implementation partner.
MOU with She’s Included
A second memorandum of understanding, with She’s Included, which convenes Nigeria’s national conference on women’s financial and economic opportunity, makes gender and financial inclusion a formal governance commitment.
The KYComm white paper
Goose FL publishes the full technical and policy design, showing how federated learning satisfies the automation mandate and data-sovereignty law at the same time: structural compliance, not procedural exception, with validated benchmarks and a defined pilot.
MOU with NITDA
Goose FL and Nigeria’s National Information Technology Development Agency sign an MOU in Toronto, anchoring the deployment on sovereign infrastructure under Nigerian jurisdiction, and committing to financial inclusion as the design goal.
Submission to the CBN Regulatory Sandbox
Goose FL submits the KYComm pilot design to the Central Bank of Nigeria’s Regulatory Sandbox: collaborative federated AML detection across participating institutions, a supervisory dashboard for regulators, and validation of the ≥70% false-positive-reduction target against live legacy systems.
If the sandbox validates, the pilot follows
The sandbox review decides the next milestone. If the ≥70% target validates against live legacy systems, the pilot becomes the working blueprint for shared compliance infrastructure anywhere data-protection law and detection mandates collide.
What we expect the sandbox will prove
Six months of testing is built to answer one question with evidence rather than argument: can institutions learn from each other’s patterns without moving a single record, and does that make detection both better and fairer?
-
Outcome 170% fewer false positives or more
Detection improves against the systems institutions already run
Success is a false-positive reduction of 70% or more measured against each participating institution’s legacy system, with accuracy above 90%, AUC above 0.92 and F1 above 0.88. A reduction under 40%, or accuracy under 80%, ends the test rather than extending it.
-
Outcome 2Under 75% of the privacy budget spent
The privacy budget holds while the model still learns
Differential privacy is capped at 8.0 epsilon over 30 days, with a 0.5 daily ceiling per institution and a delta of 1e-5. Success is clearing the accuracy thresholds while using less than 75% of that budget, so there is room left for the rounds that follow.
-
Outcome 3Parity gap under 0.1
Fairness is measured, not asserted
Demographic parity difference has to stay under 0.1, with 0.2 named as the failure line. A model that is accurate overall and wrong for one population is not a working model, and the test says so in numbers.
-
Outcome 4Complaints under 5% of the test population
Consumer protection holds, and the harm is named
Complaints have to stay under 5% of the test population and no customer’s account may be frozen in error. The people that high rates of false positives harm are the point of the exercise.
-
Outcome 5Four institutions active or more
A network, not a single institution’s model
At least four participating institutions have to be active for the result to mean anything, across commercial, non-interest and digital banks. Federated learning gains come from the network, so a test that only ever proves one bank’s model has not tested the thing that matters.
-
Outcome 6Supervisory view validated
Regulators get a working view of the model
The supervisory dashboard has to be validated by the Central Bank of Nigeria and the Nigeria Financial Intelligence Unit, suspicious transaction report quality has to improve measurably, and the evidence pack has to be usable in Nigeria’s 2027 FATF mutual evaluation.
-
Outcome 7Erasure that reaches the model
Unlearning is exercised, not promised
Erasure requests under section 34 of the Nigeria Data Protection Act have to be served by machine unlearning and evidenced round by round, not answered with a policy statement. That is the obligation shared models most often fail quietly.
