Adam M.

I build production LLM systems for industries where a wrong answer costs real money — insurance, finance, law, property. Expert-Vetted (the badge Upwork reserves for its top 1%), 100% Job Success across 70+ projects and six years. Not prototypes. Not proof-of-concepts. Production systems that non-technical people depend on every day. WHAT I DO Two specialisms: 1. Document intelligence — extracting structured, validated data from messy real-world documents (scanned PDFs, hand-filled forms, appraisals, payslips, 200-page policy manuals) at accuracy a business can put its name to. 2. Agentic LLM systems — multi-agent and multi-step workflows where AI makes real decisions: schema-validated, cross-checked by a second model, observable at every step, with a human in the right place by design. The common thread: a manual process that took days becomes an automated pipeline that takes minutes — with the evidence to prove it didn't get worse. THE RECORD • Lead engineer on a production AI underwriting platform, sold through a policy-software vendor serving 18 insurance carriers: 99.55% precision on discount and surcharge rules, 96.14% overall — policy review went from days to 3–5 minutes • 12M+ UK planning documents (120M+ pages) extracted and embedded in under 48 hours, at 65–75% below the GPT-4 baseline cost • Payslip extraction for a UK mortgage broker: 52% fewer processing errors, 80% less manual review, 15% faster mortgage approvals • 700+ UK Local Plans summarised weekly, fully automated, at 75% lower LLM cost through context caching and batch APIs • A consumer-fintech assistant where every financial figure is computed by deterministic code, not the model — zero made-up numbers by construction • A private-equity deal-screening agent that completes the firm's own investment scorecard end-to-end, backtested against realised fund returns The 99% figures are maintained properties, not launch statistics: golden-sample regression suites replay real production cases on every change. HOW I WORK I default to the simplest implementation that works and iterate with evidence. Strong prompting before fine-tuning; an API call before custom infrastructure; a benchmark before an architecture decision. Non-negotiables on every pipeline: • Pydantic schema enforcement — if it doesn't validate, it doesn't pass • Dual-model verification for high-stakes data • Full observability (LangSmith / Langfuse) — nothing is a black box • A gold-standard test set built early, regression-tested on every change Communication: daily written async updates and a weekly Loom walkthrough. WHAT CLIENTS SAY "His Loom updates were the highlight of my week! He's great at communicating complex AI concepts to non-technical people and keeps you in the loop every step of the way." — Chris Barnes, Co-Founder, Gains App "Adam is an absolute powerhouse of an LLM Engineer. He has first class communication skills which make working with him an absolute pleasure." — Sammie Ellard-King, Founder, Gains App "A great balance of personality, professionalism and a deep knowledge of the AI space. He's a strategic thinker who considers the bigger picture." — Tom Story, PlannrAI TECHNICAL STACK • Orchestration: LangChain, LangGraph, LangSmith, Langfuse, Claude Code / agent skills, MCP • Models: Claude, GPT, Gemini, Llama — via Bedrock, Vertex and direct APIs • Infrastructure: AWS (ML Specialty certified), Docker, Kubernetes, Postgres + pgvector, FastAPI • Core ML: Python, PyTorch, pandas, scikit-learn • Validation: Pydantic, structured outputs, dual-LLM verification, LLM-as-judge test suites CURRENTLY TAKING ON • Enterprise document processing — financial, insurance, legal, mortgage, property • Agentic workflows that have to be right, not just impressive • LLM cost and accuracy work on existing systems • Production pipeline architecture and technical leadership Expert-Vetted · 100% Job Success · $400K+ earned · 5,750+ hours · 70+ projects