Language, speech & machine intelligence

Language models in the loop

Drafting, summarizing and checking, under human gates.

Models draft emails, summaries and reviews from the record. A person approves every outward step. Each model run receives a packet that says exactly what it may read and write.

Examples of our work

  1. Owner-facing drafts

    Used a harness that selects the right voice and facts for an owner-facing email, memo or report, then hands the draft to a person.

  2. Adversarial reading

    Ran a falsifier pass on proposals using the smallest model able to find a flaw, escalating only when it could not.