Engineering

How the systems are built.

Short notes on production AI systems I have shipped: what the problem actually was, what the design settled into, and what broke along the way. Written for engineers.

Architecture note

Shipped · Lira Intelligence

Letting a customer plug their own tools into an AI agent

A customer wants the AI agent to do things inside their systems, not just answer questions about them. That means running their code paths from our process, which is a security problem before it is an AI problem. This is the design that resulted.

  • TypeScript
  • Fastify
  • MCP 2025-11-25
  • AWS KMS
  • DynamoDB
  • OAuth 2.0

Architecture note

Shipped · Lira Intelligence

Retrieval does not throw when it is wrong

A marketing page outranked a bank's own policy document because it used more of the customer's words. Separately, one filter clause hid an entire knowledge base. Neither raised an error. Both produced fluent, confident answers.

  • Qdrant
  • TypeScript
  • RAG
  • Hybrid search
  • Evaluation

Engineering note

Open source · evalgate

Safety metrics get no tolerance

Quality metrics drift and that is fine. But "forbidden tool calls only went up four percent, which is inside tolerance" is not a sentence anyone should be able to ship behind. Here is how the gate is built, and what it deliberately does not cover.

  • TypeScript
  • LLM evaluation
  • CI
  • Prompt injection
  • RAG