31 posts
Post #05 ended with a list of half-built features and a promise to ship them. This is the delivery receipt: compliance export, idempotency hardening, and the systems work that made them production-ready.
At-least-once delivery guarantees every message is processed — not exactly once. This post traces the three idempotency layers in the Sentinel-L7 Axiom pipeline and the specific failure modes each one defends against.
An LLM pipeline that produces no exceptions is not necessarily a pipeline that works. This post covers four signals for scoring the quality of compliance audit narratives produced by Sentinel-L7.
The previous post listed what EventHorizon needed before it could run in production. This one closes the list — persistent resume tokens, a multi-stage Dockerfile, a dependency-aware health check, and k3s manifests. Along the way, a silent data-loss bug that the test suite missed entirely.
The first version of Sentinel-L7's producer guard was binary: above the threshold it stopped, below it ran full speed. This post covers why that fails under real load and what graduated backpressure looks like in practice.
Sentinel-L7 runs two long-lived worker processes on Redis Streams. When a worker crashes mid-message, that message stays stuck in the PEL indefinitely. This post covers why the dedicated reclaimer daemon was replaced with embedded XAUTOCLAIM recovery in every worker.
Why providing a deliberately flawed starting point is the fastest way to kill ambiguity and surface requirements.
EventHorizon is feature-complete as a learning vehicle. The interesting question is which of its patterns survive the move to a real production environment, and which need to be reworked entirely.
I started this project with two AI assistants and a sync problem. I'm finishing it with one assistant and three documents that act as the project's durable memory. Here's what changed.
A speculative look at what Sentinel L7 is pointing toward — the features that are half-built, the decisions still open, and what building this thing has actually taught me.
Importing a server module should not bind a port. It should not connect to MongoDB. It should not start a metrics interval. Once your tests can't import your server without ten unrelated things happening, you've already broken your own contract.
I evaluated Valibot. I picked Zod. I lost half a day to a UUID validator that was technically more correct than the previous version. Here is what I'd tell my past self about validation libraries.
The backend Jest strategy, frontend Vitest setup, two bugs found while writing tests, and an honest account of what the app still can't do.
The dashboard said everything was fine. Stats updated. Connection dot was green. Nineteen thousand events had moved through the pipeline and the live feed was a frozen screenshot.
There are parts of this system I have no automated tests for, and that is a deliberate design decision. The trick is being honest about which parts those are and why.
AMQP prefetch is one integer in one config file. It is also the difference between competing consumers and one worker eating the entire queue while the others sit idle.
Integrating a Python anomaly sidecar, hitting Gemini quota limits within minutes, and why a pluggable driver abstraction is not over-engineering.
How I almost shipped duplicate documents into MongoDB, and why I now think of unique indexes as load-bearing infrastructure rather than schema decoration.
On building a policy knowledge base for compliance analysis — and the silent failure that made it look like everything was working when nothing was.
Why the worker republishes failed messages instead of nacking them with requeue=true, and how a single boolean flag is the difference between a healthy pipeline and a frozen one.
There are exactly seven steps in EventHorizon's graceful shutdown, in exactly one order, and every reordering loses data in a specific way.
Why I named the four stages of the pipeline before I wrote any code, what naming bought me, and what backflow between them would have cost.
How conflict detection, a status state machine, and a single role-aware endpoint handle the core scheduling logic in RhizoBook.
How Prisma's include silently exposed sensitive fields on a public endpoint, and how explicit select fixed it at the query level.
Fingerprints, embeddings, and the exact-timestamp bug that gave me a 0% cache hit rate until I stopped being clever and started being dumb.
How I structured the Next.js frontend with two route groups, wired up NextAuth JWT sessions, and replaced a fragile build-time env var with a server-side rewrite proxy.
Why I designed Sentinel L7 around three separate processes, Redis Streams, and a rule-based fallback that fires more than I expected.
Why I chose NestJS over plain Express, how I structured domains as modules, and the schema decisions that shaped the whole app.
From identifying a recurring problem in healthcare SaaS to designing a solution I actually want to use.
Crunch time in Santa's workshop
Why I'm starting to write more and consume less — and what this site is for.
no posts with that tag.