Ajin Baby · AI Platform & Cloud Infrastructure Architect
Making cloud run itself,
without losing control of it.
Designing AI-enabled cloud platforms — from agentic infrastructure and MCP servers to self-healing systems — for safety-critical and regulated industries.
I ship this in production at a global aviation-software company. The book, research, and code on this site are all free.
The book
Self-Healing Infrastructure
Building Autonomous Cloud Systems with AI
Most writing about AI and infrastructure stops at the demo. This book starts on the day the model is confidently wrong at 3 a.m. and an auditor asks who authorised the action it took.
Research
Preprints on reliability engineering for AI-enabled infrastructure, each with runnable companion code. They're self-published and not peer reviewed.
-
From Incident to Failing Build: Incident-Derived Regression Testing for AI-Agent Guardrails
A corpus of ten publicly documented AI-agent production failures, normalised into runnable reproductions behind a CI gate, with a measured ablation showing which guards are load-bearing and what configuration drift costs.
-
Stateless Protocol, Stateful Problem: Durability, Retry, and Retention Policy for MCP Tasks
The MCP Tasks extension specifies a lifecycle and leaves the operational policy open. A measured study of durability, poll cost, failure classification, and a client/server coupling that silently cancels live work.
-
Error Budgets for Autonomy: SLO-Driven Authority Management for Autonomous AI Operators
Transplanting SRE error-budget semantics onto agent authority: a blast-radius-weighted budget and an asymmetric authority ladder, with a measured simulation study.
-
Stable by Design: A Control-Theoretic Account of AI-Driven Self-Healing Remediation Loops
Self-healing systems fail as loops, not decisions. A control-theoretic model of AI remediation, its instability modes, and the constructs that settle them — measured.
Latest field notes
-
Claude Code Hooks: Guardrails That Don't Depend on the Model Remembering
CLAUDE.md is advice the model can reinterpret. Hooks run every time. Four working hooks, the exit-code gotcha that fails open, and where hooks stop.
-
LiteLLM 1.107.0: The Key Was Deleted in May. Nobody Revoked It.
The malicious release was live for two hours. Getting there took five months, an agent platform holding an owner's token, and a relaxed release gate.
-
How to Write a Postmortem Someone Will Actually Read
A practical template for blameless postmortems: when to write one, what each section is for, how to word it, and a small linter that catches the usual mistakes.
Tools I build in the open
All repositories → Upstream, I've also had a fix merged into Neo4j's official MCP server.
Who's writing this
I'm Ajin Baby: 15 years in cloud infrastructure and platform engineering, now a Staff Software Engineer working on AI platforms. Along the way I co-founded two companies and grew one from $500 in capital to ~$500K ARR across 110+ countries. IEEE member; MS in AI in progress. More about me →