Alexander – Python, Machine learning, LLM

protein R&D start-up (TUM spin-out)

• Co-led the ground-up restructuring of the company with its founders.
• Moved ProtXplain from a product-first to a concierge-driven model, in which internal tooling and products were built around customer demand.
• Helped build the MVP.
• Led the platform architecture.
• Built evidence-grounded research planning.
• Built ontology-backed exploration of alternative routes.
• Implemented human approval gates before compute.
• Built auditable end-to-end workflows.

Europe's largest OTC energy-trading venue, where utilities and traders request and close quotes for power, gas and other energy products. I worked as the sole ML consultant on the platform's data and AI capabilities and on building the company's in-house applied AI team.

• Built ML models for client- and recipient-side liquidity estimation.
• Built an interpretable trade-likelihood recommender for requests for quote (RFQs), replacing an earlier RAG pipeline misused as a classifier with tree-based models, TreeSHAP explanations, and counterfactual what-if analysis.
• Presented the business case for dedicated AI roles to the C-suite on multiple occasions and helped build the in-house applied AI team.
• Ran technical hiring for product data scientist and ML engineering roles.
• Provided ongoing advisory on applied AI roles, ML infrastructure, and model deployment.

Back-office automation for an early-stage venture capital fund: a document-intelligence pipeline that reads capital calls, distribution notices and fund reports, extracts the key fields and routes each document for processing, with human review where it matters. I led the engagement end to end, from scoping to delivery.

• Built a document-intelligence pipeline with OCR + LLM extraction and confidence-gated human review, chosen over vision-language models after benchmarking on the client's historical filings.
• Worked within the client's preferred workflow tool (n8n), with complex steps moved into separate cloud services.
• Achieved about 98% extraction success on capital calls and over 92% automated routing accuracy.
• Ensured 100% human verification of payment-critical fields, with a full audit trail.
• Cut processing time from hours to minutes per task.

German subsidy-management firm

• Led the engagement.
• Set the automation roadmap for the firm’s document-heavy subsidy workflows.
• Laid the groundwork through rapid prototyping.
• Prototyped an OCR → classification → ownership clustering → field-extraction pipeline for sensitive client documents with human-in-the-loop verification.

Industrial research lab working on machine learning for communication networks and customer-facing products. I worked on communication-efficient federated learning and led a project on LLM-based simulation for synthetic data generation.

• Led the project on LLM-based customer-complaint simulation for synthetic data generation and edge-case testing.
• Developed communication-efficient federated learning methods (selective updates, adaptive masking, iterative pruning).
• Published as first author in IEEE TGCN, co-authored an IEEE PIMRC 2025 paper, and is an inventor on a US patent application.
• Co-supervised University of Bristol dissertations, interviewed candidates, and trained new hires in AWS, computer vision, NLP, and federated learning.

Dating and matchmaking products operating in 37 countries. I built the company's ML foundations and production infrastructure, and the models for payment-fraud prevention, content moderation and trust and safety.

• Designed and built the platform for deploying models to production (containerised FastAPI microservices on Kubernetes, asynchronous worker queues).
• Built payment-fraud prevention models, using explainable-AI methods to explain predictions to the payments team.
• Developed image classification for user-uploaded content to reduce customer-support workload, including a migration from TensorFlow to PyTorch.
• Built trust-and-safety classifiers for user messages: spam (in production since 2019), support-spam filtering, and exfiltration attempts.