AI that answers to your compliance team
The useful half of AI without the data leaving your control. Local and fine-tuned models on hardware you own, personal data stripped before a prompt goes anywhere, and enterprise terms with the vendors who do not train on what you send them.
Every model worth picking, on the same line
Frontier models from OpenAI, Anthropic, Google, Microsoft and Amazon, open weight families from Meta, Mistral, DeepSeek and Alibaba, and your own fine-tunes. All of them behind one endpoint, so choosing a model is a policy change rather than a rebuild.
Open weight families run on your own hardware, or in our Dhaka data center. The frontier models are reached through the enterprise agreements above, on the routes the policy allows for that data class.
A public API is a data export
Most AI projects stall on the same three questions: where the data goes, who keeps it, and whether it trained a model somebody else owns. Answer those in writing and the project moves. Ignore them and it dies in a legal review six weeks later.
Retention
Consumer endpoints may keep your prompts for weeks. Enterprise plans let retention be turned off, but only on the right tier and only if somebody asks for it in writing.
Cross border transfer
A prompt typed in Dhaka can be processed in another jurisdiction. That transfer is a legal event, and for public sector data it is usually the one that stops the project.
Model training
On consumer terms your prompts can improve a shared model that your competitor also uses. On enterprise terms they cannot, and the agreement says so.
Audit
Regulators ask who saw what, when, and under which agreement. A route with no log is not an answer you can give, however good the model was.
Enterprise terms, not free tiers
Where a cloud model is the right tool, we call it through the agreements these vendors sell to regulated customers. No training on your data, retention disabled, and a data processing agreement behind the whole thing.
Four providers, all under contract
Microsoft Azure OpenAI
Enterprise agreement
Region pinned deployments inside your own Azure subscription. Abuse monitoring retention is disabled on approved accounts, and prompts are excluded from model training.
- No training
- Retention off
- Region pinned
Amazon Bedrock
Your own AWS account
Foundation models reached through Bedrock, so the agreement sits between you and AWS. Inputs and outputs are not stored and are not used to train any model.
- No training
- Nothing stored
- Your contract
Google Vertex AI
Enterprise terms
Model access with data residency controls, so processing stays in the region you nominate. Customer data is not used to train Google models.
- No training
- Data residency
Anthropic Claude
Direct or via Bedrock and Vertex
Claude over the provider you already contract with, or on a direct enterprise agreement. Zero retention commitment, and no training on customer traffic.
- No training
- Zero retention
Terms differ by service and by region, and they change. Before anything goes live we show you the current agreement and the retained data settings for each model we intend to call.
What crosses the border, and what does not
There are three ways a request can be answered. We pick per workload, not per preference, and you can see which one is in use from the audit log.
- 01
Stays in Bangladesh
Open weight models on your own GPUs or in our Dhaka data centre. Nothing leaves the network, so there is no transfer to assess.
- 02
Leaves masked
Personal data is replaced with tokens before the request goes out. The mapping stays with us and the real values are put back in the reply on return.
- 03
Leaves in region
Where a cloud model is genuinely the right tool and the contract allows it, processing is pinned to a region you nominate on an enterprise agreement.
Built into every deployment
- Personal data masked before routing, not after
- Model chosen by data class, cost and policy
- Per prompt audit log naming the policy that allowed it
- Region pinning wherever the contract requires it
- Bring your own cloud account, so the vendor contract is yours
- No fallback to a public endpoint, ever, without a written change
Where a workload touches citizen, patient or customer records, it runs on the first route by default. The other two exist for cases where a cloud model is genuinely better and the data is not sensitive.
AI your compliance team will sign off on
Banks, hospitals and ministries want what AI can do, without sending citizen or customer data to a public API. This is the stack we build for them.
Inside your network
Patient Fatema Akter, phone 01712 345678, was admitted with type 2 diabetes.
What the model sees
Patient [PERSON_1], phone [PHONE_1], was admitted with [CONDITION_1].
PII stripped before routing
Names, NID and phone numbers, account numbers and medical details are masked inside your network, then put back in the reply.
- gpu-node-01gpu-node-02gpu-node-03gpu-node-04
air-gapped. no outbound traffic
Models on your own hardware
Open-weight models on your GPUs, or in our Dhaka data center. Prompts and files never cross the firewall.
- gateway.btlpolicy: data class
- Local
- GPT
- Claude
- Gemini
- Titan
- Mistral
One endpoint, every model
Route by cost, speed, data class or policy. Swap a provider, or fall back to a local model, without a code change.
Then we use it on your project
- Discover
- Prototype
- Build
- Test
- Run
Fine-tuned on your own manuals and case files, and built to pass the residency, audit and access checks a public sector review asks for.
Tell us what you need to build, protect or automate.
A 30 minute call with an engineer, not a sales script. You leave with a clear next step, whether or not we work together.