Post

Every time you choose to add complexity rather than measure and improve the system you already have, you gamble with your time and your focus. The new component arrives with a tax — latency, cost, more moving parts to operate — and it pays off only when the thing underneath it was already doing its job. Without a solid foundation and good measurement, the upgrade is motion, not progress. This is a post about one instance of that gamble — adding a reranker to a search system — and how the same shape repeats somewhere less obvious: the AI agent you point at a search tool you built. In both cases the temptation is to say “the smart layer will sort it out.” It won’t. A smart layer can only refine what the layer beneath it hands it. ...

I’m excited to announce that Inboxer, a small Obsidian plugin I created for my own workflow, is now available in Obsidian through Community Plugins. As an ML engineer, I spend a lot of time in Obsidian managing research notes, and I kept running into an annoying workflow issue. The Problem I’d found there was a lot of friction when trying to add details to a note, say a project or an idea, in a structured way. I’m in the habit of keeping an ‘INBOX’ heading where I can append random thoughts, links, websites etc, and for projects a ‘TIMELINE’ heading to record events and progress. It became annoying switching to a note, finding the heading and adding a subheading whenever I had new information. Context switching kills my flow, and those small interruptions add up. ...

I’ve been investigating multi-agent systems and their applications in development workflows. After reading Anthropic’s recent guide to agent coordination patterns, I wanted to experiment beyond established patterns like Evaluator-optimizer and Orchestrator-worker workflows. For this exploration, I’ve chosen automated code review as a sufficiently complex problem domain. I’m building a multi-agent system structured according to the Viable System Model (VSM) – a cybernetics framework that provides an elegant organizational structure for complex systems. This post explores how VSM can serve as a foundation for implementing a multi-agent code review system that’s both effective and adaptive. ...

TL;DR This is a quick write-up of RetentionCast, a system I built to predict customer churn in real-time using Change Data Capture. The system processes live transaction data through Kafka and Debezium to calculate churn probability and Customer Lifetime Value. Check out the GitHub repo The Churn Problem If you’ve worked in e-commerce or SaaS, you know customer churn is the silent killer of growth. Traditional approaches to churn prediction often involve batch processing - running models overnight on yesterday’s data. But by the time you’ve identified that a customer is at risk, they might already be gone. ...

I’ve been kicking the tires on Typesense, an open-source search engine and alternative to Elasticsearch, for a personal project. So far I’m quite impressed with it. This is a small post recording how easy it was to set up vector embedding fields using a custom model. It is getting much easier to build semantic search applications. In previous work when I was setting up semantic search it was both exciting and frustrating, because the tooling wasn’t there and you had to build basically every piece yourself. It was cutting-edge, but you had to make your own cutting tools before you really saw the benefit. In my case, I had been using Elasticsearch and had to augment it with annoy, (using approximate K nearest neighbour in place of cosine distance, until Elasticsearch added vector search support). ...