Luke G.

Most AI projects fail at the data layer, not the model. I build LLM systems that run on real production data: web scraping, data extraction, RAG, AI agents, and the evals that prove the output is right. Anyone can call an LLM API now. The hard part is getting reliable data in, and knowing whether what comes out is actually correct. Recent work - A scheduled collection and classification pipeline that has run unattended in production for six months. It pulls tens of thousands of records a day from a large marketplace via API and scraping, into a structured database, filters 65% of them out deterministically before any LLM spend, then classifies and scores the rest against comparable market data. Qualifying results publish straight into a spreadsheet the client works from, so buying decisions get made on scored data rather than manual research. - Claude wired into a client's internal systems: CRM, WhatsApp and email, with agents working alongside the team in a shared channel. I scoped the permissions and connector architecture end to end. Ongoing, with more of their processes automated as we go. - A multi-agent financial modelling pipeline for a real-estate underwriting platform. Document extraction, agent orchestration with LangGraph and FastAPI, and formula-driven Excel output for scenario analysis. Taking what took analysts days or weeks to do down to minutes. What I build - Data extraction at scale. Playwright, Selenium, PyDoll, BeautifulSoup, high-concurrency async, proxy infrastructure. Pipelines with retries, checkpointing, validation and structured logging that keep running when sites change or fail. - LLM systems in production. RAG, structured extraction, agent orchestration with LangChain and LangGraph, MCP servers and Claude Connectors wiring models into the tools a business already uses. Claude, OpenAI, Gemini and open models, used programmatically. - Evaluation and reliability. Golden sets, LLM-as-judge validated against human labels, regression tests in CI. Most teams ship AI features without knowing whether quality has dropped. I build the measurement systems. - Infrastructure. Python, FastAPI, Postgres and pgvector, Docker, GCP and Azure, scheduled cloud jobs with proper error handling. Background Two years at HarperCollins owning the advertising system behind £12.5m in annual revenue, plus company-wide LLM rollouts including manuscript analysis that cut editorial review time by 70%. Three years teaching data science, so I can explain technical decisions to non-technical stakeholders. I work end to end, from scoping and architecture through delivery and iteration, with clear communication throughout.