The Silent Failure Is the Most Dangerous One: Fail-Closed in CommitBrief
On a partial config, one CommitBrief command silently turned off secret scanning. Silent failures I closed July to October, and the rule: fail closed.
Systems Architecture · Engineering Research · AI Systems Engineering
I am a software engineer and have been building software since 2008. My core is the architecture, scalability and reliability of high-traffic distributed systems in production; my current direction is AI systems engineering. Here I share the reasoning behind decisions, the lessons drawn from real projects, and measurements anyone can reproduce.
With 10 million dead rows, a partial index sustained 11,537 claims per second; the same table without one managed 7 (Postgres 17, Docker, Apple M4 Pro).
Staff Engineer Since 2008 (18+ years)
New here? Begin with these five; I chose them by hand.
The first request after a deploy across seven PHP frameworks with opcache.preload on and off: the gain, its start-up cost, and who can reach it.
The same OAuth2 + PostgreSQL API measured in Go, PHP-FPM and FrankenPHP on 1/2/4 cores under a 50/50 read-write mix: capacity per core and CPU per request.
One afternoon I hit the limits of self-review, so I wrote a tool that audits my own diff without leaving the terminal: CommitBrief.
Closing dual-write with an outbox, suppressing the at-least-once repeats it creates, and encrypting the personal data inside the event.
Cheap code production doesn't raise throughput if review and integration can't keep up. The bottleneck moves downstream — that's where the work is
Research
A claim that rests on neither a source nor a number is a preference. The record of what I looked into lives here — measurements, analyses, surveys and reviews — each stating its question, its method and how far I trust the result.
Go to the researchBlog
My latest writing on software, architecture and product development.
On a partial config, one CommitBrief command silently turned off secret scanning. Silent failures I closed July to October, and the rule: fail closed.
In CommitBrief, what counts as the 'same' finding moved from the title to a rule ID; baseline, policy and PR threads were built on that decision.
I measured CommitBrief's default models for the first time: Gemini cleared the threshold, while the OpenAI default raised too many false alarms.
New series · Free books
Three free books on one idea: reliable systems are not built from reliable parts but around the parts that fail — access, events and AI agents.
Portfolio
A selection of products and projects I build and maintain.
Founder & Developer
BabelQueue is a language-agnostic message queue standard that lets services written in different languages share the same queue without getting trapped by serialization lock-in. Instead of language-specific formats like PHP's serialize(), it defines a strict JSON envelope — frozen at schema_version 1 — that every language can read natively. It runs on Redis and RabbitMQ with no sidecar or broker plugin, with an overhead target of under 2%.
Founder & Developer
CommitBrief is a provider-agnostic Go CLI tool that performs LLM-powered local code review on git diffs. It reviews any scope — from staged/unstaged changes to a single commit to PR-style three-dot ranges — without leaving the terminal, without leaking the diff, and with customizable project rules.
Go · Cobra · go-git · Anthropic Claude
Founder & Developer
Ragmux is a self-hosted, single-binary AI gateway that sits between your application and providers like OpenAI, Anthropic, Gemini, DeepSeek and Ollama. It exposes one OpenAI-compatible API, injects RAG context from the documents you upload into every request of a project that has a store linked, and keeps provider credentials encrypted and away from the application. Users, connections, projects, documents, vectors and metrics all live in one PostgreSQL with pgvector — no Redis, no separate vector database.
Go · PostgreSQL · pgvector · ParadeDB
Expertise
The work usually starts with a role. Under each one sit the technology stacks I build it with.
I practise architecture not as a collection of patterns but as decisions made in a particular system, with a particular boundary and a particular cost.
Technology stacks
Where the provider sits, where the budget is cut, what authority a tool runs with and how output is measured: the page covers these four decisions and the seven areas behind them.
Technology stacks
Work moved onto a queue does not become durable by itself; I am a backend developer who designs the contract, the queue, the database and the server together.
Technology stacks
How a change goes live should be readable from the repository; I build that single pipeline of packaging, testing and deploying.
Technology stacks
My career and background: About Résumé
What I get asked most: how I started, what I build with, and how I choose.
5 questions
I met my first computer in 2000, an IBM my family bought. It did not take long to realise that what I really enjoyed was not using the machine but creating something with it; that feeling of making things was incredibly satisfying. My software development journey started there.
I use PHP, Go, Python, JavaScript/TypeScript and Rust; all of them are languages I write actively today. Among frameworks, Laravel and Symfony, Gin and Fiber, and the Node.js side are part of my toolbox. My focus is never the language or the framework, it is the problem to be solved.
There is no single answer, but a few parameters drive the decision: the time available for the MVP, the knowledge and experience of the rest of the team, and the specific needs of the product. If nobody on the team knows Go, that product should not be written in Go. By the same logic, I would not write PHP Laravel for a consumer/worker whose only job is to fire a single cURL or SOAP request.
This is a personal corner of the internet where I share what I know and what I have lived through, in detail. Posts, research and projects gather around systems architecture, engineering research and AI systems. It exists so that I can pass on my experience freely, with no advertising and no revenue motive behind it.
I have been using AI actively, as a driving force, since 2023. For the time-consuming grunt work such as reviewing codebases and writing code, agentic AI tools do a good job as accelerators. My current direction is AI systems engineering: I treat the model as one component of the software system that carries it. My principle is this: the machine works at the language layer, not the decision layer.