Thirteen years building things that hold.

Every team was wiring its own LLM calls — duplicated effort, reinventing the wheel per use case, slower delivery, and no shared place for provider choice, streaming, or (later) usage monitoring, guardrails and evals. So I built a shared, provider-agnostic LLM service: any provider (OpenAI and Gemini wired first-class, DeepSeek too, with Claude or any new provider/model designed to drop in), any model, streaming, and structured or unstructured responses from one interface — plus a re-entrant handler that executes tools mid-stream over OpenAI’s Assistants API. It became the backbone a whole shelf of features were built on.

A from-scratch async multi-agent orchestrator I built instead of reaching for a framework — so I had the control a customer-facing flow needs: sequential / parallel / hybrid execution, semaphore-bounded concurrency, per-task retries and timeouts, a task dependency DAG, lifecycle hooks, and cancellation. It powers a genuinely complex trip planner: specialist agents for destinations, hotels, activities, transfers, visa and insurance fan out in parallel — each grounded in RAG over real inventory and pricing — then a synthesis agent assembles the plan.

Ops ran vehicles, drivers and daily/weekly schedules by hand in spreadsheets — error-prone, hard to manage, and a direct hit to traveller experience. As we expanded across Europe, Australia and New Zealand it became a bottleneck: impossible to scale without piling on headcount, and more headcount meant more mistakes. So I built driver-roster management plus a routing engine: pickup-and-delivery with time windows (PDPTW) for a destination-management fleet — airport, intra- and inter-city, private and shared — with a build-vs-buy strategy that dispatches each batch to a SaaS optimiser or to an in-house greedy heuristic I wrote to honour 13-hour shifts, driver rest, capacity, skill matching and return-to-base. Live GPS monitoring tracks the fleet in real time.