I ship the whole system — the feature, the pipeline, and the cloud it runs on.

Litigation funding — investors financing legal claims — traditionally runs on paperwork, phone calls, and trust. AEQUIFIN's marketplace digitizes it, which means the platform must do three hard things at once: move investor money correctly, protect privileged legal documents, and satisfy European compliance law. A bug here isn't a glitch — it's someone's money or someone's case file.

Software engineer on the core platform — building new features end-to-end and maintaining existing infrastructure, from database schema migration to production rollout, on a live bilingual (EN/DE) system.

Real balances, privileged case material, GDPR obligations, and zero tolerance for stale or corrupted records.

Every release, engineers manually translated Figma designs into React code compliant with a 350+ component design system — a full-time tax across dozens of product teams. Every existing tool failed the same way: Figma's own AI scored under 30% accuracy, V0 and Bolt.new have no concept of your design system, and fine-tuned models hallucinated props that don't exist. None of them know what your components are allowed to look like.

Owned implementation end-to-end, pairing with the platform architect in an XP-style setup — they owned architectural calls, I owned the build.

Hard correctness requirement — generated code had to use only real components with real props. No budget for a labeled fine-tuning dataset. Output had to be trusted without a human reviewing every generation.

A global cable manufacturer had machines consuming energy around the clock and only end-of-month invoices to explain where the money went. The tools on the market were reactive — they reported consumption after the fact but couldn't predict a spike before it happened, so inefficiencies were discovered instead of prevented.

Built the web application end-to-end: a unified sensor-integration framework with adapters for heterogeneous IoT sensors, real-time ingestion, and the prediction layer on top.

Live factory-floor data streaming over WebSockets; power-factor and consumption dashboards that plant operators — not just engineers — could read; a sensor network that had to scale without re-architecting.

Biopsies are invasive, painful, and risky — and they're still the default step in cancer diagnosis. The Sonocure project aimed to use ultrasound waves and biomarkers for early detection and targeted treatment instead.

Built ML models and the pipeline around them — training, integration into the Angular/Node application, and CI with Jenkins.

Medical-grade ambition on research-project timelines; models had to adapt ultrasound parameters per patient, not one-size-fits-all.