Nuwan M.

I build Claude and LLM systems (RAG, agents, extraction pipelines) that you can check. If your product depends on an LLM or on messy records, I build the tests, matching rules and audit trail that show when the output is wrong, before your users find out. What I have built - A production Claude pipeline for a US health-technology client (since April 2026), which I build and maintain. It runs behind about 1,500 automated tests, and every fix ships with a regression test. - A clinical-trial registry pipeline (Python, PostgreSQL): a reproducible database built from 23 registries. Each run fingerprints its code and sources, and gates pin the verdict counts so a dropped test fails the build. Five Claude extraction stages are scored against labelled sets fixed in advance. Every extracted value must quote its source, and a prompt change ships only if it beats the measured noise. - Bioinformatics engineering for a UK genetics company (Python libraries, Nextflow pipelines, backend APIs), and before that two years of Nextflow pipelines in Docker and Singularity plus HPC administration at the Sri Lanka Institute of Biotechnology. Work I take on - LLM evaluation: test sets, regression harnesses, LLM-as-judge checks measured against human labels, citation and fact checks. - Independent review of an AI system or AI-built app before go-live: a written report of what breaks, with tests that prove it. - Extraction, matching and reconciliation: messy records matched to a reference with confidence scores, a human review queue for the uncertain ones, and counts that reconcile. No invented values. - Full-stack development: Next.js and React front ends, FastAPI and Node.js backends, PostgreSQL and Supabase, authentication and third-party API integration, built test-first and deployed with Docker. - Bioinformatics pipelines and biology-domain AI evaluation (genomics, GATK, Nextflow). How I work I agree the checks with you before I write the code, so we both know what done means. You get small, reviewable changes, a test for every fix, and a short note on what changed and why. Stack: Python, TypeScript, SQL, FastAPI, PostgreSQL, Supabase, Node.js, Next.js, Docker, Nextflow, AWS, Anthropic and OpenAI APIs, Claude Code. Background: BSc Biotechnology (University of Colombo), BSc Information Technology (University of Colombo School of Computing), MSc Computer Science in progress (ML, NLP). Springer Nature publication. Based in Sri Lanka (GMT+5:30), available full time.