Shipped systems, redacted names.

After iOS 14, the brand had lost roughly half of its Facebook ad attribution. Acquisition cost was climbing with no diagnostic. On the payments side, failed recurring charges on SEPA Direct Debit were leaking subscribers every month.

Implemented the Facebook Conversions API server-side, deduplicated with the existing pixel, hashed PII at the edge, and added drift monitoring on top. Re-architected the subscription model to support multiple concurrent subscriptions per user. Re-implemented the Stripe flow with smarter retries to recover failed Direct Debit payments.

A US healthcare group with operations across many facilities had grown its internal platform organically. Auth was fragmented across third-party tools, the Django backend was slowing under load, and the operations team had no unified view into either patient operations or internal workflows.

Built a centralized authorization layer on Azure API Management so every third-party system sits behind a single secure gateway with fine-grained role and access control. Led a team of 4 engineers on the migration of the legacy Django backend to FastAPI, with automated testing, structured logging, and production monitoring in place from the start.

An internal service running CSV-triggered workflows through a Camunda business-process orchestrator had become a bottleneck. The backend was a Django monolith with very little observability, and a separate legacy Django service depended on it with almost no documentation.

Tech-led the migration to FastAPI with a team of 3 engineers and a Product Owner, including a redesign of the Camunda integration. Exposed backend metrics over SNMP and Prometheus exporters so the platform team could observe the service in production. Built a React frontend for the internal users. In parallel, kept the undocumented legacy Django service running as the sole developer on it.

Production decision systems often need to express uncertainty over outcomes, which standard ML toolkits don't really support. This shows up most in sequential and causal decisions in e-commerce and finance.

Bayesian probabilistic models for production datasets in e-commerce and finance, with simulation pipelines wired into backend systems so inference drives decisions instead of living in notebooks. Workloads containerized with Docker and shipped through GitHub Actions CI/CD.