Five years building systems that can’t afford to fail. Started at Tejarat Bank — fraud detection for 5 million customers, the kind of job where a missed transaction becomes a news story. Then London: real-time APIs, production ML pipelines, sub-100ms inference.
Now at Anthropic via Alignerr, stress-testing how AI coding agents handle the work that actually breaks real software — ambiguity, silent scope reduction, overconfident claims. Adversarial, high-signal.
Drawn to problems where latency, scale, and correctness all matter at once. Fraud at 300k transactions a day. Distributed state. Inference pipelines. Anywhere “works on my laptop” doesn’t cut it.
Available for full-time backend and ML engineering roles in London. No sponsorship needed until January 2027.
Design real-world software engineering tasks to evaluate AI coding agents (Claude Code) — backend features, refactors, debugging, test coverage across Python codebases. Score models across 14 engineering-behavior dimensions and construct adversarial scenarios to surface failure modes like overconfidence and silent scope reduction. Qualitative feedback feeds directly into model training.
Owned the production fraud detection API end to end — FastAPI, Redis, Celery, RabbitMQ, XGBoost, Kubernetes, MLflow. Shipped a loan default model to AWS Lambda with drift-triggered retraining. Built an AI-powered job matching platform with GPT-4o-mini and sentence-transformers. Plus a Django consulting platform that lifted Lighthouse performance by 45%.
Contracted to fix database performance issues on a 1,600-student platform. Composite PostgreSQL indexing cut query time from 8s to under 200ms. Shipped FastAPI microservices with JWT/RBAC on Kubernetes with auto-scaling, Qdrant semantic search, GitLab CI/CD, Prometheus + Grafana observability.
Built automated fraud detection in Python and Oracle SQL, processing 10,000+ daily transactions. Rule-based anomaly detection and statistical scoring cut manual investigation time by 30%. Tuned Oracle queries and rewrote stored procedures for 40% faster compliance reporting. Integrated Qdrant for fraud pattern similarity, reducing false positives by 25%.
End-to-end fraud detection service processing 300,000+ transactions daily at sub-100ms inference. XGBoost + Isolation Forest ensemble with SMOTE on a highly imbalanced dataset (0.1% fraud rate). Full MLOps pipeline with MLflow tracking, drift-triggered retraining, Kubernetes deployment.
CV-to-JD semantic matching platform using GPT-4o-mini for skill extraction and sentence-transformers for similarity scoring. Fuzzy matching with 80+ skill synonym mappings. Deployed full-stack on Railway and Vercel with a live demo.
Brought in on contract to fix critical database performance at KodjoEnglish (1,600+ students). Composite PostgreSQL indexing cut query time from 8 seconds to under 200ms — a 97% improvement across 50,000+ daily queries. Shipped Kubernetes-deployed FastAPI microservices with Qdrant for semantic content search.
An async Python bot that delivers music on demand to 1,000+ Telegram users. A user sends a track query, the bot fetches the audio from the upstream source and returns the file. The engineering centre is an SQLite-backed cache layer: repeated queries skip the external API entirely, cutting calls by 60%. Concurrent request handling through asyncio with connection pooling and graceful rate-limit handling. FastAPI management layer, fully Dockerized.