I build the boring-critical stuff — pipelines that move 50M+ records a day without waking anyone up at 3am — and the new stuff: LLM systems with retrieval, function calling, and evals that actually hold up in production.
- 🏦 Deutsche Bank — event-driven ingestion at scale, JVM/G1GC tuning on a 24-node Hazelcast cluster, ~$500K infra savings
- 🛡️ OSTO Cybersecurity — multi-tenant data isolation at the Postgres layer, concurrent collector pipelines, mTLS-secured high-throughput event ingestion
- 🤖 Now — agentic systems: retrieval routing, structured outputs, LLM-as-judge evals
|
Java 25 Virtual Threads · Kafka fan-out embeddings · pgvector HNSW + Redis RAG 
|
Deterministic scoring + LLM enrichment · FastAPI · PostgreSQL
|
|
DDPM diffusion for tabular data · Polars streaming beyond RAM
|
Data-integrity fixes merged into a 37k-star finance terminal 
|