Senior AI Engineer | Agentic Systems & LLM Platforms

Production assistant that turns a surveyor’s dictation into a finished expert report (customer in production). Speech-to-text (ElevenLabs Scribe v2) with case-specific key terms, A/B-tested; dictation cleanup with a content-loss guard; facts only from the dictation, open points flagged instead of invented; RAG style archive of past reports with hybrid retrieval (vector + full-text, Reciprocal Rank Fusion, EU embeddings); layout from templates (DOCX/PDF), never from the LLM; cost tables extracted from offer PDFs with sums computed in code; critique loop feeding the expert’s corrections into regeneration; realtime voice interface (OpenAI Realtime) with tool results interpreted by Claude.

Designed and built, essentially single-handedly, a multi-tenant SaaS AI company operating system with an operator agent and five department agents (marketing, sales, content, design studio, development). Own agent runtime (earlier: LangGraph) with nested delegation, multi-agent orchestration, 400+ typed tools across 40+ systems, RAG on pgvector with hybrid retrieval, governance with human-in-the-loop approvals, sandboxed coding agents for website deployment, a voice and document AI report assistant for building surveyors, and an engineering system for AI-written code with 1,000+ architecture contract tests and LLM-as-judge quality gates. Platform complete and handed over to operations in October 2026.

Built CompanyOS, the agent platform that preceded Linkworld. Agent workflows with LangChain, from 2024 with LangGraph: supervised build and deployment workflows with approval steps (human-in-the-loop), agent orchestration as a workflow graph with state and handoffs between steps.