TWZNBooking Q4 2026
Practice02 / 04
02AI

Cortex

AI integrated into the systems you already run.

AI woven into the systems you already run.

  • RAG over your own data
  • Tool-using agents
  • Evals & cost budgets
01Thesis

Demos Are Easy. Operations Are Hard.

Every AI demo works. Far fewer survive contact with a shop floor, a support queue, or a Tuesday afternoon outage. Cortex is AI integration for production: models grounded in your real schema and business logic, wrapped in validation and fallbacks, and placed where they change daily decisions — not in a chatbot no one opens. We've run this in live retail operations where the people asking the questions are staff behind a counter, not data scientists.

02Capability
01 / 03

Ask the Database Anything

On Intouch, retail staff ask questions in plain English — “Who's our best customer?”, “Show me sales by product type” — and get answers in under two seconds. Behind it: a schema-aware pipeline where Gemini generates validated, SELECT-only SQL, executes it, and a second model formats the result — returning charts, not paragraphs, when the data warrants it. Conversations persist, so “show me their purchase history” knows who their refers to. When generation fails, hand-crafted fallback queries answer common patterns instead of returning an error.

Fig.AI query pipeline
PipelineAnswer
Staff
Asks a question
“Who is our most loyal customer?”
Stage 01
SQL generation
schema-aware · with history
Stage 02
Read-only check
one SELECT, validated
Execute
postgres · read-only role
Stage 03
Format + chart
context-aware answer
Output
Answer + chart
“Sarah Chen — £12,400 across 23 orders”
5 events · play, or hover a node to trace its path
Full conversation persistence — follow-up questions have context.
Conversational BI pipeline — IntouchView case study
03Capability
02 / 03

Automation That Writes Itself

A workflow copilot turns intent into infrastructure. Describe it — “send a thank-you email two days after every purchase over £1,000” — and Claude generates the full trigger–delay–condition–action chain on a visual canvas staff can edit. Underneath, a real execution engine: 20+ domain events, step-based runs with context chaining, template variable resolution, and a full run history. The AI writes the workflow; the engine makes it real.

The events those workflows react to come from systems wired by Relay

Fig.Workflow engine
ExecutionDesign
Assist
AI copilot
claude · English → workflow
Visual builder
Canvas
react flow · drag & drop
Workflow v12
triggers, conditions, steps
App event
order.shipped
API returns at once
Fire-and-forget
Runner
supabase · async
Send email
1 of 20+ handlers
Delay
wait 2 days
Webhook
context chained
6 events · play, or hover a node to trace its path
Drag-and-drop design with fire-and-forget async execution.
Workflow execution engine — IntouchView case study
04Capability
03 / 03

The Right Model for Each Job

No single model is best at everything, and no single vendor should be a critical dependency. Cortex architectures are multi-model by default: on Intouch, Gemini handles SQL generation (latency, structured output), Claude powers the workflow copilot (multi-step reasoning), and OpenAI drives campaign content. Each is chosen for the task — and each is replaceable without touching the product.

Fig.Model routing
Router3 tasks · 3 models
Task
SQL generation
low latency · structured output
Task
Workflow copilot
multi-step reasoning
Task
Campaign content
creative generation
Cortex router
Task → model
picks per task: latency, reasoning, cost, output shape
Google · swappable
Gemini 2.0 Flash
Anthropic · swappable
Claude
OpenAI · swappable
GPT
6 events · play, or hover a node to trace its path
One router, many models — each task dispatched to the model best suited to it, and any model can be swapped without touching the product.
Multi-model AI strategy — IntouchView case study
05By the numbers
  • question to answer
  • AI providers, 0 lock-in
  • events triggering AI workflows
  • only, always validated
07Engagement
Options02

Ways to start

Every engagement starts with a conversation. Fixed-price entry points, then work scoped to the system it has to carry.
  • 01Start here

    Production Pilot

    £5,000fixed · four weeks

    One workflow, your real data, live in four weeks. We pick the single highest-leverage place AI changes a daily decision, build it into your real systems with validation and fallbacks, and ship it to the people who'll actually use it.

  • 02The full engagement

    AI Integration

    Scoped per project

    Multi-model architectures, conversational BI, workflow copilots — scoped against your systems and priced per project. The pilot usually tells us both what this looks like.

Not sure where you fit? Tell us about the system and we'll point you at the right start.

Talk to us
09Contact
London--:--:--

Have systems that should be smarter?

If your team still answers questions by exporting CSVs, or runs processes a model could handle, that's where we start. Tell us what you run today.

  • AI Integration
  • LLM Pipelines
  • Copilots
  • Multi-Model

What happens next

  1. 01ReplyWe reply within 48 hours
  2. 02CallA 30-minute call to understand the problem
  3. 03ProposalA written proposal with scope and price