Writing
Patterns from building with AI. Each one links to something that runs.
Evaluating LLM Output in Production: Validate, Repair, ScrubChecking LLM output is easy. The hard part is what to do when it fails: a repair that fixes only the failing sentence, and a scrubber that cuts it.Two towns named Calheta: debugging an LLM verification pipelineA real receipt, refused three times for three reasons. Three bugs at three layers, and why designing the failure direction first made being wrong survivable.There are characters you cannot seeSome Unicode characters render as nothing, and a model reads them as text. How I made prompt injection useless on my travel site. Not resisted. Useless.178 reports in one afternoon: what a publish burst does to an LLM pipelineA traveller published 178 trip reports at once. My site had 9 before that. The queue between the trigger and the model, and what held.I pay an LLM to approve bad reviewsEvery trip report goes through a model before readers see it. The most important line in that prompt is not about catching bad content.You just write. The places find themselves.You write the trip the way you would tell a friend. Minutes later the places you mentioned are on a map. Where the model’s job ends.The librarian pattern: how I keep my AI coding assistant from breaking my appOne index file, one doc per feature flow, and a 40-line bash hook. How I keep an AI coding assistant from breaking my app.Context-Lens: a serverless, open-source MCP server for AI document understandingHow I built an MCP server for semantic search over local files and GitHub repositories.Unlock Bedrock InvokeInlineAgent API's hidden potential with Multi-Agent OrchestratorUsing Bedrock inline agents inside a multi-agent orchestrator.Beyond auto-replies: building an AI-powered e-commerce support systemA customer support system where several AI agents share the work, built on Multi-Agent Orchestrator (now Agent Squad).Introducing CloudFront Hosting ToolkitA CLI that builds a full front-end deployment pipeline on S3 and CloudFront in two commands.How DAZN uses AWS Step Functions to orchestrate event-based video streaming at scaleEvent-driven orchestration for live video streaming.Prototypes
Small tools I built for a real need at home. Open to try, not products.
I built a tool with a 3D car to help my son learn to driveBefore his driving lessons, my son couldn't picture what happens inside a car: what is connected to what. With my coding assistant I built a tool around a 3D car. Then came parking, the motorway, the blind spot, and driving it yourself.The Search Grammar Pattern: Natural Language Search with LLMsHow I built natural language search for my movie app with one small LLM call, and the pattern behind it: describe the offer by its dimensions, not its rows.