Engineering Notes
Field notes from production AI and ERP work
Long-form guides on ERPNext development, agentic AI architecture, AI security and AI integration in business workflows. Written from systems that shipped and stayed running — including the parts that went wrong.
AI security engineer
What Is an AI Security Engineer? Role & Skills (2026)
An AI security engineer secures the parts of a system that reason, retrieve and act. Here is what the role covers, how it differs from AppSec and ML engineering, and what the work looks like day to day.
Read the guide →
prompt injection
Prompt Injection Attacks: How They Work and How to Defend
Prompt injection is not a bug in your prompt. It is what happens when instructions and data share one channel. Here is how the attacks work and the defences that contain them.
Read the guide →
OWASP Top 10 for LLM applications
OWASP Top 10 for LLM Applications (2025), Explained
All ten 2025 OWASP LLM risks, each with a realistic example, a test you can run this week and the control that closes it.
Read the guide →
AI red teaming
AI Red Teaming: A Step-by-Step Methodology for LLM Apps
Red teaming an LLM application is not a pile of jailbreak prompts. It is scoping, threat modelling, a curated attack corpus, measured results and regression tests.
Read the guide →
MCP security
MCP Security: How to Secure Model Context Protocol Servers
Every MCP server you connect is code you run and text your model reads. Here is the trust model, the attacks that exploit it, and the controls that keep an MCP deployment from becoming a backdoor.
Read the guide →
AI agent security
AI Agent Security: Excessive Agency and Least Privilege
A chatbot that is tricked says something wrong. An agent that is tricked does something wrong. Here is how to size permissions, gates and sandboxes to the damage an agent can actually do.
Read the guide →
RAG security
RAG Security: Securing Vector Databases and RAG Pipelines
A RAG pipeline is a data system with an LLM on the end. Here is where it leaks, where it gets poisoned, and the retrieval-time controls that actually enforce who sees what.
Read the guide →
LLM guardrails
LLM Guardrails: Input and Output Controls That Actually Hold
A guardrail is not a classifier you bolt on. It is a stack of input, context, tool and output controls, and the deterministic ones do most of the work.
Read the guide →
how to become an AI security engineer
How to Become an AI Security Engineer: A Practical Roadmap
There is no degree for this yet. Here is the six-stage path I would follow again — from web pentesting to building LLM apps, breaking them, defending them and showing your work.
Read the guide →
ERPNext security
ERPNext Security: A Hardening Checklist for Frappe
An ERP holds your money, your customers and your payroll. Here is the hardening checklist I work through on self-hosted ERPNext, from permission levels to Redis ports.
Read the guide →
AI integration in business workflows
AI Integration in Business Workflows: From Pilot to Measurable ROI
The blocker is almost never the model. It is process selection, system access and the absence of a baseline you can measure against.
Read the guide →
AI security
AI Security: Threat Modelling, Red Teaming and Guardrails for LLM Systems
Prompt injection is not a bug you patch, it is a property of the architecture. Here is the threat model, the red team method, and the controls that survive contact with attackers.
Read the guide →
agentic AI development
Agentic AI Development: Architecture Patterns for Production Multi-Agent Systems
A demo agent takes an afternoon. A production agent takes tool design, an evaluation harness and a budget ceiling. Here is the difference, in architecture terms.
Read the guide →
ERPNext development
ERPNext Development: Custom Apps, Workflows & Integrations That Survive Upgrades
Most ERPNext projects fail at the second upgrade, not the first deployment. Here is the architecture that keeps customisation, automation and integrations maintainable.
Read the guide →