At MPIRIQ (Munich, remote), I built a clinical-trial matching engine that parses eligibility criteria with LLM function calling and scores matches with cited evidence, cutting manual screening effort by ~80%. I re-architected trial and cohort matching into a hybrid system where SQL handles structured criteria and the LLM only handles free text, lowering costs by ~2.7x.
I also took a cohort query platform from 0 to 1, shrinking cohort creation from hours to minutes, built a clinical document pipeline with agentic MedDRA / ICD-10-CM coding, and raised LLM extraction throughput by ~5x. I deployed an event-driven notification service on GCP with Pub/Sub, Kubernetes and Terraform, and mentored 2 junior AI researchers.
Before that, at Gravitas AI (London, remote), I built an agentic voice bot with ~1s end-to-end latency, saved ~$150K/year with a chatbot architecture redesign, and delivered an agentic legal chatbot over 50,000+ documents that cut search latency from 15s to 5s. I raised the ICMR medical chatbot’s accuracy from 65% to 97%, fine-tuned Llama, Mistral and Gemma with QLoRA and Unsloth (~25% better accuracy, ~70% less training compute), and mentored 4 interns and 2 engineers.

