LOOP71
Replay the loop
ENع
Start a project

AI, generative AI and machine learning

AI, built
properly.

We design, build and run AI that does real work: products, agents, search over your knowledge, forecasting and vision. We already run it inside our own platforms.

Request an AI buildTry the live model

01Signal

Your data is noise
until something learns it.

Orders, documents, sensor readings, conversations. Millions of points with no shape yet.

02A base model

A base model
finds the patterns.

A few layers, trained on your examples. It classifies, summarises and answers simple questions. Good, and cheap to run.

03A deep model

A deep model
reasons.

Dozens of layers, retrieval over your documents, tools it can call, a plan it checks before acting. This is where agents live.

04Yours

Then it takes
your shape.

Your rules, your brand voice, your approvals. We build at the depth your problem needs, and run it where your data is allowed to live.

Request an AI buildTry the live model

Describe your idea. A tiny model routes it.

This is a real machine-learning model, deliberately small: a naive Bayes text classifier trained in your browser on 120 example sentences. Nothing you type leaves this page.

Take this into a brief

Made with our own local model

Name your industry.
Watch an idea appear.

A model running on this machine writes one AI idea for your sector, live, in a sentence. Nothing you type is stored or sent anywhere else. When no local model is running, you see one of our thirty illustrative briefs instead, and the label says so.

Model depth

From a base model
to something very deep.

Drag the slider. A small model answers simple questions. A deep one reasons across documents, plans multi-step work, uses tools, checks itself and remembers your context. We build at the depth your problem needs, and no deeper than it should be.

4 layers0 connections drawn

  • Answers simple questions
  • Writes and summarises
  • Reads your documents with citations
  • Reasons across many sources at once
  • Plans multi-step work and uses tools
  • Checks its own answers before you see them
  • Remembers your context across sessions
  • Works in English and Arabic, with voice and vision

Illustration of depth and capability, not a diagram of a specific model.

What we build with AI.

01

LLM products

Assistants, copilots and writing tools built on large language models, wired into your product.

02

Agents and automation

Agents that take actions in your tools, with clear limits, logs and approvals.

03

Search your knowledge (RAG)

Answers grounded in your documents and data, with citations people can check.

04

Machine learning and forecasting

Models that predict, rank and detect, scored honestly against what really happens.

05

Computer vision

Cameras and images understood on the edge or in the cloud, privacy first.

06

Voice and conversation

Voice receptionists, call summaries and multilingual conversation, including Arabic.

07

Generative media

Images, video and motion for campaigns, produced with local and cloud models.

08

Evaluation and safety

Test sets, red-teaming, permissions and human review designed in from day one.

Already running inside our own platforms.

Forecasting and machine learning

Pathora

Official economic and market data becomes indices, scenarios and forecasts, each one traceable to a receipt and scored in the open.

Open the case
Product film, made in our local AI video studio. The figures shown are from one logged research run (27 Sep 2026), not a forecast or accuracy claim.
  • 238klines of code
  • 85data feeds registered
  • 7composite indices
  • 10in-house forecasters
  • 880+automated tests passing

TypeScript · React · Node.js · SQLite · Python · custom canvas charts

Computer vision at the edge

Tessera

Privacy-first video safety for yards and docks: existing cameras, faces blurred, every event checked by a person.

Open the case
Product film, made in our local AI video studio. AI-generated scenes; the alerts are illustrative, not real Tessera detections.
  • 197klines of code
  • 589safety checks catalogued
  • 31hazard categories
  • 18jurisdiction rulepacks
  • 2,400+automated tests

C# · .NET · Python · TypeScript · Apache-2.0 vision models

Retrieval and evidence AI

Veyr

A research question becomes a cited evidence report: two independent assessors, a search for counter-evidence and a human with the final word.

Open the case
Product film, made in our local AI video studio. Interface shown with demonstration data; research software, not medical advice.
  • 83klines of code
  • 1,428automated tests
  • 11research data connectors
  • 21database tables
  • 9checkpointed pipeline stages

Python · FastAPI · PostgreSQL · Redis · Celery · local open-weight models

A local vision model on a PC

CxS

A private gaming companion that runs on the player's own PC, answers from sources they saved and looks at the screen only when asked.

Open the case
  • 33klines of code
  • 420+automated tests
  • 35database tables
  • 6game launchers detected
  • 0cloud calls needed to answer

Python · JavaScript · local vision model · zero third-party Python packages

How we ship AI you can trust.

  1. Define the task

    One job, measured.

  2. Connect the data

    With permissions and provenance.

  3. Evaluate

    Test sets before launch.

  4. Guardrails

    Limits, logs and approvals.

  5. Human review

    People decide what matters.

  6. Monitor

    Scored against reality.

Request an AI build