Occasional notes on the things I build and the problems around them — media provenance, secure delivery, and running backend systems in production.
The EU AI Act's Article 50 and California's SB 942 took effect on the same day. A look at what is actually deployed for media provenance — generation-time watermarks, C2PA metadata, fingerprinting — and the gap the in...
Read →Everything you send to a model is free text, and free text leaks. data_redactor now ships an opt-in RubyLLM integration that scrubs PII and secrets from every outbound request — prompts, tool definitions and tool resu...
Read →Feasibility research done before a hackathon: what breaks when you try to recover a watermark from a screenshot of a rendered 3D model, and why the failed experiments turned out to be the thing worth having.
Read →Two weeks, a stack I'd never shipped, and an AI doing most of the typing. What I learned was less about the tech than about how to point it.
Read →API keys, IBANs, credit cards and national IDs leak into logs, indices and audit trails. data_redactor is a Ruby gem with a C extension that finds and replaces them — 88 built-in patterns, scanned in a single pass.
Read →Europeana's search endpoint and record endpoint don't return the same schema with fields omitted — they return a flat index document versus a knowledge graph. What that difference means when you build on the API.
Read →Two fields sit quietly in every Europeana record: contentTier and metadataTier. What they mean, how they're calculated, and what a given combination tells you about a record's quality — gathered into one place.
Read →Propagating a trace context from a FastAPI gateway into an asynchronous Temporal workflow turned into a rabbit hole. The fix was deleting the custom interceptor and using the one that already existed.
Read →An hour lost to Nginx logs for a failure that never reached Nginx. The default 4MB gRPC message limit was rejecting uploads inside the local process — a lesson in auditing every hop in the request lifecycle.
Read →Split every 500 characters and you cut sentences in half. semantic_chunker is the first Ruby gem for true embedding-based semantic splitting — grouping sentences by meaning rather than by character count.
Read →Ruby's Array hits a performance wall on insertion and deletion near the head of a large list. DLinked brings the doubly-linked list to Ruby with an idiomatic interface and O(1) operations at both ends.
Read →