I fell in love with neural networks in 2016. I worked in robotics and multi-agent reinforcement learning, and then focused on the systems around models.
Multi-agent RL and emergent communication My academic focus (BSc and MSc at FIAS / Goethe University Frankfurt), and still the thing I find most interesting: what happens when independent learners have to coordinate, and what kind of communication protocols fall out of that pressure. Peer-reviewed publication at AAMAS.
Robotics and representation learning Simulation-based experiments with PyRep and CoppeliaSim — manipulation tasks, learned visual representations, and the question of how much structure you need to impose versus how much the agent can discover. Convolutional architectures throughout.
End-to-end ML platform, agrochemical domain Built and architected a full platform: data ingestion, feature and experiment management, training, deployment, monitoring. Less about any single model than about making the path from raw data to a served prediction reproducible for a team that wasn't going to babysit it.
Agentic RAG systems Retrieval pipelines with agents doing the planning and decomposition rather than a fixed chain — including in regulated environments where provenance and auditability aren't optional. Currently interested in context aggregation as the real bottleneck for coding agents.
BSc and MSc with strong focus on image recognition and reinforcement learning at FIAS. Worked as ML Engineer, Data Engineer, and led a team impplementing an E2E ML platform in pharma. Currently in financial services consulting, working on AI solutions in credit, payments, and securities.
- I'm learning rust!
Timtody (from Perl’s motto “TMTOWTDI”) is basically the idea that in Perl there are a million ways to solve the same problem. I don't know any Perl, luckily.