The Accelra blog
Notes from
doing the work
Products, AI, data, and infrastructure. The decisions we made, what happened next, and what we learned along the way.
Projected monthly savings
- From · billed
- $798.62/month
- To · projected
- ~$206/month
Full-month invoice confirmation pending
Featured / Infrastructure / 12 min read
Cutting the infrastructure bill
From $798.62 billed to about $206 a month projected: roughly 74% in expected infrastructure savings for PokerInk. The changes are deployed; the full-month invoice is still pending. Here is what we changed.
Read the storyInfrastructure
1 articleSecurity
3 articles13 min read
Secure Foundations: HIPAA-Grade AWS with Terragrunt
AWS gives you the compliance rubric for free. Passing it is the work. The four jobs of a secure AWS foundation — prevent, detect, fix, alert — plus the hard-won lessons from a real HIPAA healthcare build.
14 min read
Software Developers: Attacks like this are going to happen again, how to limit the blast radius
The LiteLLM payload was 332 lines of Python. It stole SSH keys, AWS credentials, API tokens, and crypto wallets from 500,000 machines. Here's the actual code, what it would find on your machine, and the developer playbook to defend against it.
12 min read
The Trivy Supply Chain Attack: RCA
An AI agent compromised Trivy's 32,000-star repo by exploiting a misconfigured GitHub Actions workflow. A different AI agent could have caught it. The full root cause analysis of the TeamPCP campaign that cascaded through five ecosystems.
Agentic Engineering
10 articles20 min read
Most Teams Are Skipping the Hard Part of Agentic AI
The real path from chat to agents is not smarter models. It is better human-agent workflows, better repo context, and tighter operational boundaries. The 5-stage roadmap most teams should follow.
14 min read
Your AI Has Root Access. Now What?
Your AI agent has API keys, cloud access, browser sessions, and GitHub repos. Rules-based security fails because context resets every session. Here's the architectural approach that actually sticks.
22 min read
How to Build Your Own AI Teammate
Everyone says 'just use AI' but nobody explains how to onboard one as a real teammate. The practical decisions behind identity, hardware, access, memory architecture, secrets management, and the operational patterns that make a 24/7 AI employee actually work.
15 min read
I Gave an AI Its Own Laptop and Made It Employee #1
Five weeks ago, I bought a Mac mini, created a company email address, gave it a GitHub account, and told it to start working. Here's what I learned running an AI as a full teammate — the failures, the security scares, and the multi-model protocol that changed everything.
18 min read
The Road, Not the Cars
AI tools made code cheap. But code without infrastructure is cars without roads. Here's what actually breaks when AI-accelerated teams outrun their deployment pipeline.
11 min read
How We Operate PokerInk with Agentic Engineering
The operating cadence behind PokerInk: ship, verify, challenge assumptions, and close loops before bad ideas get expensive.
18 min read
Push to Main. Skip the PR.
We push directly to main — no feature branches, no PRs. Instead, a validation environment with Playwright E2E tests and a pipeline that actually catches bugs. Part 1 of the Agentic Engineering Series.
13 min read
Your Software Has a Pulse
What happens when your software stops being a static artifact and starts being a living system that monitors itself, responds to change, and evolves? Part 2 of the Agentic Engineering Series.
16 min read
Your AI Has Opinions. Most of Them Are Wrong.
AI agents optimize for helpfulness, not doubt. Here's the multi-model challenge framework that catches confident wrong plans before they cost you weeks. Part 3 of the Agentic Engineering Series.
8 min read
The Death of the PR
Pull requests review the one thing that matters least. After 15 years shipping software infrastructure, we've been verifying the wrong artifact the whole time.
These articles describe particular projects and operating decisions. Client delivery approaches are agreed for each engagement.