Currently shipping
Tightening the eval harness on my self-hosted RAG prototype: adding retrieval-quality tests for entity-heavy edge cases (policy numbers, acronyms, abbreviations) that pure semantic search misses. The same citation-first pattern powers the chat at the top of this site, so the harness work feeds back into both.
On the side: Riff, a voice-first AI music production partner. Right now I'm wiring the brief schema into an audio engine compiler. Hard problem, fun problem.
Reading + thinking about
Anthropic's tool-use patterns and the latest writeups on agentic loops over LLMs. Reverse-engineering what Claude Code does well as an agent harness. It's the cleanest one I've used and I want to apply the same patterns to multi-step back-office tasks.
Opinion of the week: hybrid retrieval (semantic + keyword) is underrated. Most "RAG isn't working" stories I read are pure-semantic teams who haven't realized keyword-aware retrieval would solve their entity-recall problem.
Looking for
Remote Automation Engineer, Integration Engineer, or AI Engineering roles where production-grade ops experience matters. Strong fit: a small-to-mid team building internal tooling or AI-augmented back-office workflows. Comfortable across the stack, from low-code/no-code platforms (ProcessMaker is the same family as Zapier and n8n) through Python services to AI infrastructure. Banking domain is a plus, not a requirement.
If you're hiring or have a tip, reach out. Or scroll up and ask the assistant what I built last week.
Around the desk
Accra, Ghana. Coffee in the morning, headphones in the afternoon. Claude Code as a daily driver.