Curious.
I build AI agents and test them for made-up answers and tricks that leak data. How often does your agent make things up, or leak data when someone tricks it? I'll measure it on your own data, then fix what fails. Each build comes with a test report you can rerun.
- Prompt injection and agent security testing. I check what your agent does when an email, document or web page carries hidden instructions, which tools and data it can reach, and whether risky actions like sending or deleting wait for your approval. You get the fixes, and a check that they don't make the agent worse at its job. Proof: my prompt injection guard made my agent worse.
- Testing the agent you already have. I build a test set from your own questions and documents, measure how often the agent makes things up or gets tricked, fix what fails, and leave you tests to rerun after every model or prompt change. Proof: a sample check-up of a support bot.
- Knowledge agents over your documents. Every answer cites its page, and the agent says so when your documents don't hold the answer. Proof: 188 graded answers over product datasheets.
- MCP servers. I build remote MCP servers with access tiers, rate limits and tools that do only their job, and I watch them for signs of attack. Proof: mcplytics.
- Safe writes to business systems. A gate checks what an agent writes before it lands in your books or records. Proof: tally-aiagent.
Write to [email protected], or summon me at