Blog

Field notes from building compliance AI

AML monitoring, identity verification, entity resolution, and agentic systems, written from the systems I build and run for African markets. Figures are drawn from live production data wherever one of my systems is involved.

Start here

The argument the rest of the writing builds on.

AML14 min read

Why Rule-Based AML Systems Fail (and What to Build Instead)

Thresholds teach launderers where the line is, the alert economics never work, and the labels are biased. A practitioner's tour of the four structural failures, sketched out, plus a migration path to ML that regulators can live with.

March 202514 min read

The systems, in depth

Three deep dives into software I built, maintain, and run in production: AfricaPEP and SENTINEL.

Building an Open-Source PEP Database for All 54 African Countries

Case Study

February 202515 min read

Building an Open-Source PEP Database for All 54 African Countries

The engineering behind AfricaPEP, the open-source PEP screening platform I build and maintain: 48,000+ profiles in production, two-stage name matching with published precision and recall, AML-safe merge thresholds, and the bugs that taught me the most.

Entity Resolution at Scale: Matching Names Across Languages, Scripts, and Borders

NLP

October 202416 min read

Entity Resolution at Scale: Matching Names Across Languages, Scripts, and Borders

Real algorithm scores on hard African name pairs, blocking for 1.2 billion candidate pairs, merge thresholds set by the cost of being wrong, and Fellegi-Sunter probabilistic linkage: how AfricaPEP matches names in production.

Building Autonomous Compliance Agents with LangGraph

AI Agents

September 202415 min read

Building Autonomous Compliance Agents with LangGraph

Six Gemini agents on a 10-node LangGraph state machine, running live: parallel fan-out and the state-merge bug it caused, Redis-to-WebSocket streaming, deterministic floors under every LLM path, and an evaluation dashboard that grades the system honestly.

Field notes

Shorter reads on the wider problem space.

Deepfake Detection in Production KYC: What the Benchmarks Don't Tell You

Computer Vision

January 202512 min read

Deepfake Detection in Production KYC: What the Benchmarks Don't Tell You

The attacks users actually try (mostly not deepfakes), cascade architecture over ensembles, the skin-tone problem in liveness detection with the published evidence, and the adaptation loop that gives single-signal defenses a shelf life of weeks.

Graph Databases for Beneficial Ownership: Mapping UBO Networks with Neo4j

KYB

December 202413 min read

Graph Databases for Beneficial Ownership: Mapping UBO Networks with Neo4j

Effective ownership is a product over paths, cycles are the domain rather than an edge case, and the graph should be a rebuildable projection. The honest case for Neo4j in KYB, with the Cypher-vs-SQL comparison done fairly.