Engineering
Engineering
How we build things that have to keep running. Trade-offs stated rather than hidden, failure modes named, and a section in each piece for what we would do differently.
These are written by the people who operate the systems they describe. The applied side - the same engineering done for clients rather than for our own product - sits with the company at neob.ai.
Evaluating voice agent quality
Word error rate tells you about the transcriber. It tells you almost nothing about whether the caller got what they rang for.
Monitoring production AI agents
An agent can return 200 on every request and still be failing every conversation. Request-level monitoring cannot see that, which is why agent observability is a different discipline.
Building multi-channel agent systems
Four channels is not four agents. If it becomes four agents, it happened gradually, and it happened inside the adapters.
Designing AI agent memory
The interesting question is not how an agent remembers. It is what it is entitled to remember, and where that record lives when someone asks you to delete it.
Streaming and non-streaming AI architectures
Streaming is not a performance setting. It is an architectural commitment: once output has left the building, you cannot revise it.
Managing context in long conversations
A long conversation is not a long prompt. Treating it as one is how agents end up confidently contradicting something they said twenty turns ago.
Designing reliable agent tool calls
Most agent reliability problems that look like model problems are tool design problems. The model is asked to compensate for an interface that should not have been that shape.
Building low-latency voice agents
Latency in a voice agent is a pipeline property. You do not fix it by choosing a faster model - you fix it by making sure the stages stop waiting for each other.