AI stack · Models
The model is a choice, not a religion.
We pick the model per job, then settle where it runs, what it may see, and who signs the answer. A second model is cheaper than a wrong one.
Founded 2022 · Founder in the Salesforce MVP Hall of Fame · Dubai HQ · Kochi studio
- 01Models · Claude, Gemini, OpenAI, Bedrock
- 02Agentic systems · Agentforce, CrewAI, ADK, A2A
- 03Agentic memory · Neo4j, Cognee, IndyKite
- 04Automation and voice · n8n, Retell
- 05Deployment and governance · identity, evals, policy, audit
On this layer
What we put in.
The honest no: if the work is a software factory with a named delivery date, we are the wrong partner.
Claude (Anthropic)
Deploy Claude enterprise-wide via API, AWS Bedrock, or Azure.
Google Gemini
Gemini for Workspace and Vertex AI, implemented and governed.
Jev (TypeSafe)
System One model for the small decisions. Pairs with the model that writes.
Sarvam AI
Hears and reads 22 Indian languages, speaks 10. Written back to Salesforce.
OpenAI
GPT models and the Agents SDK. Azure OpenAI when you need it.
AWS Bedrock
Private multi-model AI inside your VPC. AWS HIPAA BAAs and SOC 2 available.
Choosing
One model for everything, or a model per job.
One model for everything
A model per job
- Picked from a demo
- Picked from your own examples
- Every request at the top price
- Easy work routed to a cheaper model
- Hosting decided later
- Region and data terms named before the pilot
Where it shows up
The work this layer does.
Use cases where these tools carry the job. Each one names the stack.
Claude in Slack
A shared teammate in the channel on one named job, with approvals in the thread.
Legal document review
Contract review, regulatory research, and compliance drafting.
Contract management
Lifecycle review with graph-backed clause memory.
Compliance monitoring
Policy monitoring with an audit trail your risk team can read.
AI Customer Service Transformation
Complex cases stay with humans; Claude handles the rest.
AI IT Support Transformation
Gemini plus ADK for ticket triage and knowledge lookup.
Content creation
Drafting and localisation inside Workspace.
Health awareness
Multilingual coaching with Workspace-grade identity.
Cultural heritage guide
Grounded answers over collections and archives.
Who picks the model
Three desks, three different models.
Typical choices, each made from the work rather than the demo.
Example
Head of data, Riyadh consumer lender
Today: Every team called its favourite model. Security cannot say which region holds customer data.
After: Two or three candidates run your own cases, and the provider data terms are named in writing first.
Banking and financeExample
Service lead, Dubai property manager
Today: Every tenant message goes to the most expensive model, including the ones that only need a queue.
After: Easy routing goes to a cheaper model or to Jev. The frontier model writes the replies that need writing.
Property managementExample
Operations lead, Kochi NBFC
Today: Borrowers answer in Malayalam, and an English-first model hears noise.
After: Sarvam hears the call in the borrower’s language, and the outcome lands on the loan record.
Sarvam AI
How we pick
From your examples to a model.
Step 1
Your examples
Real cases from your desk, with the right answers marked.
Step 2
Two or three candidates
Each runs the same set.
Step 3
Pass, cost, region
The model that passes at a cost you accept, where your data may sit.
Every change
The set runs again
A model update has to pass before it goes live.
Questions
What buyers ask us.
- Which model should we use?
- The one that passes your evaluation set at a cost you accept. We test two or three candidates on your own examples before anyone commits, and we often route easy work to a cheaper model.
- Can the data stay in our cloud or region?
- Usually. Claude and other models run through AWS Bedrock or Azure in the regions they support, and Gemini runs on Vertex. We name the hosting region and the provider data terms in writing before a pilot starts.
- Where does Jev fit?
- Jev answers typed questions with a probability and never writes text. We pair it with the model that writes, so routing and scoring decisions are cheap, fast and logged.
- Where does Sarvam fit?
- When your customers speak Indian languages. Sarvam hears 22 of them, speaks 10 and reads documents in all 22. We connect it to Salesforce and keep Jev on the closed decisions. It has no Arabic, and its managed service runs in India.
The rest of the stack
One layer is not a system.
The plane this layer lives in, and the other four layers.
Runtime plane
Where the model runs, and what calls it.
Agentic systems
How agents plan, call tools, and talk to each other. Frameworks and protocols, not a single vendor lock-in.
Agentic memory
What the agent remembers: graphs, identity, and long-term context so answers stay grounded.
Automation and voice
Workflows and voice agents that connect models to Salesforce and the rest of the stack.
Deployment and governance
The work that turns a pilot into a production system: identity, evaluations, policy and audit.
Before you pick a partner
Read these first. Then call anyone.
Our buyer guides. They apply whether or not you hire us.
Choosing an AI partner in the UAE
Five kinds of partner compared, the PDPL, DIFC, Arabic and WhatsApp questions, and red flags.
Choosing an AI implementation partner
Evaluation criteria, deployment models, governance fluency, and red flags.
AI partner evaluation framework
A five-part scorecard: depth, governance, integration, security, delivery.
Questions to ask an AI partner
Architecture, security, governance, pricing, and who is accountable.
AI partner vs in-house team
Hiring timelines, real costs, and dependency risk, scenario by scenario.
AI pilot to production
Which of the four ways your pilot stalled, and the sequence that moves it.
The brief
Send the situation.
Name the process. We tell you which layer is missing. The assessment is free: a short read of the org and a written note. If you go ahead, the first build week is a fixed AED/SAR 14,900 (about USD 4,050), credited in full when you sign the build within 60 days.