A single in-country control point for every AI call: authentication, PII redaction, model routing, token metering and complete audit logging. Lets a regulated bank swap or upgrade models without rewriting applications — and keeps prompts, context and logs inside the sovereign cloud boundary.
Retrieval-augmented generation engineered for audit: hybrid retrieval and reranking, grounded prompts with source citations, guardrails and PII handling, immutable audit logs, and human-in-the-loop for high-impact answers. Designed to explain itself to a regulator, not just to demo well.
Moved a ~100,000-row dataset from JSON to TOON (Token-Oriented Object Notation) inside a production LLM pipeline and measured roughly 60% fewer tokens with a ~4% accuracy improvement on data-retrieval tasks — a lossless round-trip applied only at the prompt boundary.
Designed and built a backend that sustained 2,500,000 concurrent connections. The kind of load engineering — backend, frontend and database working as one system — that now underpins how I architect AI platforms that have to stay up under real production pressure.
Want specifics — architecture, numbers, constraints — for a particular build? The write-ups live in my writing, and the canonical profile is at ercin.info. For anything confidential, reach me via the links below.