Research that ships into products.

Research areas

• Autonomous agents with a whole job to do

  An agent that runs an entire job without being prompted step by step: find local businesses, read their reviews and audit their web presence, decide what's worth fixing, build the fix, set up a working phone receptionist and write to the owner. We study how such an agent checks its own work before acting, holds back what it isn't sure about with a reason a person can read, and pauses by itself when a source it depends on goes down.

  Used in Tendly

• One persona, many surfaces

  A single agent with one memory, one knowledge base and one set of tools, which can appear as an avatar, a cursor on a web page, a chat thread or a voice on the phone. We study how to keep its behaviour and state consistent as a conversation moves between them.

  Used in Servly

• Grounded answers in Arabic dialects

  Answers that come from an organisation's own documents, in the dialect the person speaks: Modern Standard Arabic, Emirati, Saudi, Qatari, Kuwaiti, Bahraini, Omani, Egyptian and Levantine. Retrieval, intent matching and answer checks are built for Arabic first.

  Used in Servly

• Agents that act inside interfaces

  Agents that point, highlight, click and fill on real web pages while explaining what they are doing. They never fill in passwords, one-time codes, card numbers, IBANs or national IDs. Pay, submit and delete actions turn into point-and-highlight so the person makes the final click.

  Used in Servly copilot, Pairly

• Acting for one person, safely

  An assistant that reads someone's email and documents must treat that content as data, never as instructions. Background work that handles inbox content can't browse the web or rewrite its own skills, and nothing leaves on the person's behalf (an email, an invitation, a booking) until they approve the exact text, recipients and account. Each company the person runs keeps its own memory and files, walled off from the others.

  Used in Keel

• Real-time voice

  Speech in, speech out, quickly enough to hold a natural conversation. That covers web voice paired with on-screen actions and phone calls that take orders, bookings and messages, with every call summarised afterwards.

  Used in Tendly, Pairly

• Handing off between AI and people

  Knowing when an agent should step aside. The agent hands over when its confidence is low or when the person asks for a human, and passes along a summary so the human agent doesn't start from zero.

  Used in Pairly

• Evaluation and knowledge gaps

  Evaluation harnesses that test agents against real questions before release, and gap detection that shows which questions the knowledge base still can't answer, so content teams know what to write next.

  Used in Servly