Who said what, and was it right
A speculative project on AI assurance for court interpreting, and two problems underneath it: putting each turn against the right speaker, and keeping an eager LLM judge honest about which flagged errors are real.
Rory Scott
Rory Scott on applied AI and data science – agentic systems, the guardrails that let them touch real infrastructure, and keeping machine-generated work traceable.
Latest posts
Terraform from my phone
Two days directing Claude from my phone while it stood up an AWS organisation, a permissions-bounded identity for itself, and a working AI demo on top. Mostly notes on the guardrails that made that a sensible thing to do.
An escalation ladder for LLM judgement
Batched mid-tier models make the calls and adversarial verifiers try to break them. I read the two places they disagreed – notes on a cheap architecture for judgement work at scale.
Analytical Writing as Applied AI: a Method, Built From a Phone in Japan
A repeatable rubric-scored method for analytical blog posts – and what it means that I built the whole thing mobile-only with Claude Code, on holiday in Japan, and actually trusted it.
What National Anthems Quietly Stopped Singing
I scored all 195 national anthems on how warlike they are, twice – the full official text against the verses people actually sing. Most countries have quietly retired their bloodthirstiest lines, God and guns turn out to repel, and the monarch is the rarest theme of all.
Password-Protected Drafts on a Static Site: StatiCrypt, Jekyll, and Honest Security Testing
How I added AES-encrypted preview posts to a GitHub Pages blog — with a 13-vector security test to prove it actually holds up.
Rapid Vibe Coding on Full Fat Models
TL;DR
IATI Traceability, Upgraded: Graph Database + MCP
Revisiting international aid traceability using a Neo4j graph database and Claude MCP — a fresh take on a problem I first tackled at the UK Department for International Development.
iTunes Wrapped: Elasticsearch + XML + Claude MCP
A day-build that parses an iTunes XML library, indexes it into Elasticsearch via Docker, and queries it using Claude MCP — a personal 'Spotify Wrapped' built with open tools.
A ‘Day-Build’ with Cursor and Typescript
Building a quantum-art visualisation in a single day using Cursor and TypeScript — an experiment in agentic front-end development and what it feels like to ship fast.