Intelligence

Intelligence is definable blocks: a block takes communication in and writes structure out - extraction, qualification & scoring, summarization, sentiment & intent, evaluation, or a custom block you define. Today blocks run on call outcomes; where the system is headed is placement anywhere in the pipeline - attached to a single step, a conversation, or the whole stream (in build). Every insight lands back in communication state, where the next block - and the next agent - can read it. That loop is the point. Because outcomes accumulate in one durable system of record, the memory can become a model: outcomes feed a dataset builder, the dataset drives RL simulation, and the result is a per-org SLM (~1B) served at your inference endpoint. You keep the trained model. A competitor arriving later starts at zero on the one input that cannot be bought, scraped, or copied: your own conversation history.

​

The primitives

​

On the spine

Blocks read interactions, outcomes, and context; they never touch the audio path. What a block produces is written back to communication state - onto the outcome it analyzed or into the Context Graph - so downstream blocks and agents read it as ordinary state, not as a side channel. Each insight written emits an event, like any other change on the spine.

​

What stays hidden

The training pipeline internals: how the dataset builder samples and labels outcomes, the RL simulation environment, and the fine-tuning stack behind the per-org SLM. You define blocks, ratify evaluations, and call the inference endpoint; how the model gets from your outcomes to that endpoint is not a surface you configure.

​

Status

• Definable blocks: extraction, scoring, and summarization run on call outcomes in production today via Post-call data; placement on arbitrary steps and streams is in build.
• Evaluations: conversation-level evals callable today through SuperDialog evals.
• Per-org SLM training (datasets, RL simulation, trained models): early access.

​

Go deeper