I build production LLM systems for industries where wrong answers cost money — insurance, finance, law, property. Recent systems have:

• ⚡ Cut insurance underwriting review from days to 3–5 minutes at 99.55% precision
• 📄 Processed 12M+ documents in 48 hours on Kubernetes spot instances
• 📊 Screened private-equity deals end-to-end against the firm's own scorecard, backtested against realised returns
• 🛡️ Built a consumer-fintech assistant designed to keep responses on the information side of the FCA advice boundary — zero hallucinated financial figures by construction

Client names are under NDA; the systems, numbers and engineering are real. Each repo has the full designed write-up:

• AI underwriting platform — the carrier's actual manual applied at 99.55% precision; review in minutes, not days
• A financial assistant that never makes up a number — every figure computed by deterministic code; 43 tools, zero hallucinated numbers
• Testing AI that never answers the same way twice — golden-sample regression, LLM-as-judge, test suites 48× faster
• When the document is the attacker — the five-layer prompt-injection defence adopted as a production platform's security spec
• Enterprise payslip data extraction — dual-LLM verification for mortgage approvals: 52% fewer errors, 80% less manual review
• National-scale document pipeline — 12M docs / 120M pages embedded in 48 hours, zero-cost self-hosted embeddings
• UK Local Plan policy summariser — 700+ plans weekly, 75% LLM cost reduction via context caching
• Client testimonials — six years of reviews, 100% Job Success

Further case studies (multi-carrier platform architecture, agentic engineering enablement, PE deal screening) available on request.

• Accuracy you can audit — golden-sample regression suites, LLM-as-judge testing, human-edit-count evaluation. 99% figures that are maintained properties, not launch statistics.
• LLM security that ships — designed the five-layer prompt-injection defence a production platform adopted as its security spec, including a ~20-line pre-filter that closes a published 100% detection-evasion attack.
• Cost-conscious architecture — 65–75% LLM cost reductions through model routing, context caching, batch APIs and self-hosted embeddings.
• Agentic engineering — led team-wide AI-tooling adoption; a ten-gate pre-push pipeline runs five adversarial AI reviewers on every change for ~$25, so PRs arrive clean before a human looks.

• Orchestration: LangChain (v1), LangGraph, Langfuse, LangSmith, Claude Code / agent skills, MCP
• Models: Claude (Bedrock + Anthropic API), GPT, Gemini (Vertex), Llama
• Infra: AWS (Bedrock, Lambda, ECS, SQS, S3, RDS), Docker, Kubernetes, PostgreSQL + pgvector, FastAPI
• ML: PyTorch, scikit-learn, pandas · AWS ML Specialty certified

Expert-Vetted on Upwork (top 1%) · 100% Job Success · $400K+ earned · Upwork profile