Posts

A new model called Jev has recently taken the AI world by storm. It’s rapidly become one of the most popular models in the agentic space, and for good reason. In this post I’ll share my thoughts on why I think that is, where this really should fit in your agentic architecture. To understand Jev’s behavior, I also experimented with creating a self-improving prompt injection detector, where the detector is a Jev call. I might talk more about that in a future post, but this post is based on a reading of the docs, and on trying it out with the detector. ...

Three places this claim-triage POC would break at enterprise scale The first two posts in this series covered where I draw the line between deterministic code and LLM judgment in a claim-triage agent, and the infrastructure that makes a “pause for a human” step durable enough to survive a process restart. Those were about what the system does and how it stays reliable. This one is about what it deliberately doesn’t do yet — because a proof of concept that only shows what works is a demo, and I want this repo to signal architectural judgment, not just implementation competence. ...

Reliability is an infrastructure problem, not a prompt problem In the first post of this series, I focused on where to draw the line between deterministic code and LLM judgment. Once those boundaries were set, the next challenge was making the system actually reliable. In the context of a claim triage agent, reliability isn’t just about the model giving the right answer—it’s about the system surviving the reality of how humans actually work. ...

I’ve been trying to build an agentic app, to understand the various trade offs design considerations involved. One thing I decided early on was to focus more on the design decisions and architecture, while letting Claude Code handle the coding bit. 🏢 Business Context In the insurance industry, the triage phase of a claim—determining coverage, verifying identity, and assessing eligibility—is often a primary operational bottleneck. Traditionally, this process relies on manual review or rigid legacy rules that fail when faced with the nuance of unstructured policy documents. The business challenge is balancing high-throughput processing with the absolute necessity of accuracy; in a regulated environment, a single “hallucination” in a coverage determination can lead to significant financial leakage or regulatory non-compliance. ...

Most agentic coding tools are shifting from subscription-based pricing to token-based pricing — a change that affects enterprise CTOs, architects, platform engineers, and finance leads equally. This shift changes the economics of vibe coding and its more disciplined cousin, spec-driven development. Earlier pricing models charged per seat or per request; now it is pay-as-you-go. GitHub Copilot announced usage-based billing; Anthropic followed suit with token-based enterprise pricing. What’s Happening GitHub Copilot’s billing is built around AI credits, where an AI Credit consumption is calculated as the token usage multiplied with a “multiplier” factor. An AI credit costs $0.01, so the dollar cost of a request is essentially the AI credit consumption divided by 100. ...

I’ve been using Claude Code heavily for the past few months, and at some point I started getting uncomfortable with what it could reach. Not because it had done anything wrong — it hadn’t — but because the blast radius if something did go wrong was my entire home directory. One rogue prompt injection in a skill file or a malicious string lurking in some library’s README, and Claude dutifully rm -rfs the wrong thing or reads my SSH keys into a response. It can’t run sudo, sure, but it doesn’t need to. Everything that matters to me as a user lives in ~. ...

This is a personal site — somewhere to write, and somewhere people can find out who I am and what I work on. The constraint I gave myself was no hosting bill. That turned out to be a useful forcing function: it pushed everything toward clean, minimal architecture, where each component has exactly one job. The about page has a chatbot on it. That was a deliberate choice, not a novelty — if someone wants to know whether I’ve worked with Kubernetes or led a platform team, they can just ask rather than scanning bullet points. It also gave me something interesting to build. ...

I recently started using SpecKit for personal projects to understand how it handles a growing codebase. I went in with the mental model of one feature = one spec. That held up well early on; it started showing cracks as the project grew and incremental updates became harder to slot into SpecKit’s constructs. SpecKit is an open source toolkit from GitHub that aims to provide a Spec Driven Development (SDD) workflow around coding agents. The core idea is to define the specification (spec) first and then drive the agentic software dev based on those specs. In theory, this should make the code more robust as the agent is following clearly defined requirements. Of course, the specs themselves need not be entirely written manually, though the requirement at its core must come from the human. Taken to its logical conclusion, the spec is treated as the source with the LLM acting as a compiler of sorts that converts the spec into the code. How well this matches reality, was what I was trying to explore. ...

For the last several months I’ve started running my own homelab on a single cheap server machine sitting on my home LAN. The primary motivation for this was to familiarise myself with running a server and multiple services on bare metal and understand how to manage such infrastructure. Infrastructure I chose a refurbished Dell Optiplex 7040 Mini PC to host the infrastructure. As multiple workloads were expected, I chose to install Proxmox VE on this device. Proxmox is a Debian derivative Linux distribution that is designed to act as a hypervisor for VMs and lightweight Linux Containers (LXCs) from a common compute and storage infrastructure. It can be used to manage a data center with multiple physical hosts, but in my setup there is only one server. ...

Most of what I write here comes from problems I’ve actually run into — designing systems that hold up under pressure, making cloud infrastructure on AWS and Azure less painful, working with data at scale, and lately, separating signal from noise in agentic workflows.