smista.ai
smista.ai
Notes on deterministic multi-model routing, traceability, and local-first AI workflows from the smista.ai team.
RSS feed
We are finally reaching the last phase of the development before the release of the first version of smista.ai.
How smista.ai routes prompts to the right model without ever asking an LLM to make the decision.
Ollama makes local AI accessible. The next step is learning when to use local models, when not to, and how they fit into developer workflows.
How smista.ai uses Rig to build a common interface for LLM providers, supporting local models, dynamic catalogues, and safer integrations.
A practical look at using Rig to build a Rust abstraction over multiple LLM providers, tools, and memory.
An AI session is not just a chat history. It is a chain of messages, routing decisions, tool calls, approvals, plans, diffs, traces, and eventually cost information. Milestone 2 was about giving all of this a proper home.
Most companies don't need fewer powerful models. They need to stop using them for everything.
Only 10 days after writing the first line of code, smista.ai has completed its first milestone: a shared core library that defines how the CLI and router communicate.
Every time I switch models, I'm not just changing a dropdown. I'm changing cost, latency, reasoning behaviour, provider UX, context, skills, and sometimes the entire workflow. smista.ai is my attempt to turn AI model selection into a routing problem: local-first, policy-driven, and built for developer workflows.