New Holland, Lincolnshire
I build things that work, then I try to break them.
I'm Nathan Graves. I make fast websites and host them properly, and I audit AI systems for the security holes that arrive with them. Same person doing both, which is usually the point.
What I actually do
Three things, properly, rather than a long list I'd have to subcontract half of.
Websites, built and looked after
Built to fit the business, not dragged out of a page builder. Fast by default, because that is the thing search engines and customers both quietly reward.
- Custom design and build
- Hosting and domains handled
- Changes without a support ticket
- No stack of monthly subscriptions
AI security auditing
LLM systems fail in ways ordinary applications don't. I look for the paths where model-controlled values reach privileged operations, and write up what I find so a developer can act on it.
- Prompt injection and tool-abuse review
- RAG and vector-store poisoning risk
- Agent permission and guardrail audits
- Findings with reproductions, not a scanner dump
Developer tooling
MCP servers, integrations and automation. Mostly built because I needed them myself first, which tends to show in how they hold up.
- Model Context Protocol servers
- API integrations and pipelines
- .NET, TypeScript and Python
- Open source where it makes sense
How it actually goes
One person, one conversation, and a straight answer. No account manager, no discovery phase that bills before anything exists.
Built here, in Lincolnshire
Every site and audit is done by me, on this desk. That is the whole operation, and it is why you get an answer the same day rather than a ticket number.
Writing
Mostly AI security and .NET. Things I hit in real work and thought were worth writing down properly.
Your Semantic Kernel agent may still be vulnerable after the CVSS 9.9 fix
The CVE was patched. The pattern behind it wasn't, and it is still sitting in a lot of .NET agent code.
Your text-to-SQL agent trusts the LLM more than you think
What happens when the only thing between a prompt and your database is a model's good judgement.
Everything else
Agent tooling, .NET, and notes on shipping software as one person.
Open source
Model Context Protocol servers, built for my own use first and then cleaned up and published. MIT licensed.
linkedin-mcpPost to LinkedIn from an agent, behind an explicit capability gate
devto-mcpDev.to drafting and analytics, with publishing deliberately withheld
meta-mcpInstagram and Facebook Page posting via the Graph API
x-mcpX posting, with rotating-refresh-token handling that survives real use
threads-mcpThreads posting on the official API
github.com/N-GravesThe rest, including Printify, Gumroad and Etsy servers
Tell me what you're dealing with
A site that's slow, an agent you're not sure you can trust in production, or something you can't find anyone else to build. I'll give you a straight answer on whether I'm the right person and what it would take — including when the answer is no.
