Hands-on AI engineering lead who takes LLM systems from a business idea to governed rollout. At Leonteq, a FINMA-regulated Swiss financial institution, I lead the AI Foundry: the internal AI platform, the deployments running on it, and the AI strategy I proposed to the CIO. The job is a delivery loop: embed with business teams, scope their requests into agent workflows, prototype fast, and carry each build through partner teams and audit review toward production go-live. Underneath sit ten-plus years of production engineering in trading and regulatory systems, alongside leading the Sophis core platform team.
Focus areas: Agent Engineering · AI Platform Engineering · RAG Systems · On-prem LLM Serving · AI Governance & Auditability · LLM Evaluation & Testing · Observability
Leading the AI Foundry, Leonteq's internal AI platform and service layer. Three delivery streams run on it: a knowledge app, agent workflows built from business requests, and agents embedded in existing internal applications. I lead all three, hands-on in Python; larger builds run with delivery partners under my project lead.
Tech lead for the client-onboarding platform, automating compliance and regulatory workflows to reduce manual effort and onboarding risk.
Leading 4 developers across Zürich and Lisbon on the firm's core trading and risk platform. Primary interface between the technical team and business stakeholders.
Implemented generic booking model for mirror booking, reducing client onboarding complexity. Built Grafana/Kibana monitoring. Key contributor to LIBOR migration.
Sophis v6.3 to Fusion 7 migration, automation tools for Trading & Treasury (C#, WPF, SQL), analytics platform for instrument pricing coverage, regulatory projects (MiFID II, SFTR, 871m).
Distributed end-of-day reporting systems, barrier monitoring, automated instrument creation (~80% manual work reduction).
Research into how LLMs deployed in Swiss financial institutions can be audited against FINMA supervisory expectations (Guidance 08/2024): SP-Benchmark, a 56-question assessment of LLM comprehension of Swiss structured products scored by dual LLM judges; per-decision audit trails from attribution graphs and token-level introspection; drift detection using sparse autoencoders; and a directional-error transcript scanner probing model conduct under adversarial pressure.
AI systems journal where topics and news are auto-collected by an agent, pushed to the repo via GitHub Actions, triggering a fresh deployment.
Self-improving AI trading system inspired by Karpathy's autoresearch. 8 agents compete on Kraken Futures perpetuals while the system improves them overnight by mutating configs, testing variants on past data, and keeping winners. Agents reflect on mistakes and build rules; when problems persist, the system writes its own code fixes.
Live near-Earth asteroid explorer built around two AI systems: a tool-calling chat agent that drives a 3D solar-system scene (selecting asteroids, changing views, running impact simulations), and an autonomous monitoring agent on a 4-hour cron using an advisor pattern (a low-cost Haiku executor that escalates higher-stakes calls to an Opus advisor), with persistent memory and email alerts. Data from NASA's NEO API.