Partnership · OK Model research partnership

The lab behind the model that runs the agents.

An operating system for AI agents is only as good as the model underneath it. Since January 2025 we have built the OK Model with AI Lab Bangladesh: they lead the research, we build the platform it has to run on. It is the model that will power the CodeFix Agentic OS at launch on 1 January 2027, and the partnership is signed through 2030.

See what we build together Ask about the Agentic OS

AI Lab Bangladesh Limited

Sector: Artificial intelligence research
Location: Banani, Dhaka, Bangladesh
Partnership: OK Model research & platform engineering
Began: January 2025

Signed through 2030

The partner

An AI lab that will tell you when you do not need AI.

AI Lab Bangladesh Limited builds artificial intelligence products for organisations that have a business objective first and a technology question second. Their proposition is plain: "Build the AI your business needs." They work from Banani and Gulshan in Dhaka, with a distributed team.

What made them the right partner is their habit of saying no. They visit a client's premises, look at the objectives, and give an honest assessment of whether AI makes sense at all, allowing that cost, culture, process maturity or the payback period may mean it does not. They are candid enough about their own field to say, "Eventually, AI will become so easy. You won't need to hire us."

That is the temperament a model programme needs. Building the OK Model is a long series of decisions about what a model should not be asked to do, and a partner who over-promises on capability would have cost us years.

Field Applied AI research & development
Based in Banani & Gulshan, Dhaka
Shape Small, senior, distributed team
Approach Honest feasibility before implementation

What the partnership covers

Model research
Training & fine-tuning
Evaluation
Inference engineering
Safety & alignment
Agent tooling
Bengali & English language
Enterprise deployment

Why a partner

What we could not do alone.

We are an engineering company. We build systems that stay up and stay correct. A model is a different discipline, and pretending otherwise would have shown up in the product.

A model is not an application

Agentic work needs a model that plans, calls tools and knows when to stop. That is research with an uncertain schedule, not a feature with a delivery date.

Enterprises will not send data away

The Agentic OS is sold on security, privacy and controlled AI operations. That rules out building the product on a model we do not control and cannot run inside a customer's boundary.

The market is bilingual

Our customers work in English and Bengali, often in the same sentence. General models are weakest exactly there, and it is not a gap you close with prompting.

Cost has to survive scale

An agent that thinks for a minute per task multiplies across a company. Inference economics had to be a research constraint from the start, not a problem discovered after launch.

What we build together

One programme, two disciplines.

AI Lab Bangladesh owns the model and the science around it. CodeFix owns the platform it runs on and the product it ships inside. The line between the two is deliberate, and the evaluation suite is where they meet.

The OK Model — AI Lab Bangladesh leads

The model that powers the Agentic OS: trained, measured and hardened for agent work rather than for conversation.

Training & fine-tuning

The data programme and the training runs behind the model, aimed at planning and tool use rather than at open-ended chat.

Tool calling & planning

Teaching the model to decompose a command, choose tools, pass structured arguments and recognise when a task is finished.

Bengali & English

A model that holds its quality across both working languages of our market, including text that switches between them mid-sentence.

Evaluation suite

The benchmark that decides whether a model version ships: real agent tasks, scored the same way every time, run before any release.

Safety & alignment

Refusal behaviour, prompt-injection resistance and the boundaries an agent must not cross when it is acting on a company's systems.

Quantisation & distillation

Smaller variants that keep the behaviour that matters, so the economics work at the scale an enterprise actually runs.

The platform — CodeFix leads

Everything the model needs in order to be a product: somewhere to run, something to act on, and a record of what it did.

Inference infrastructure

Serving the model on our own and on customer hardware, with the throughput, batching and latency budgets the product is sold on.

Agent runtime

The layer that runs many agents at once, schedules their work, coordinates them and brings the results back to one place.

Tool & connector layer

The systems an agent is allowed to touch, described in a form the model can use and an administrator can restrict.

Isolation & permissions

Per-tenant boundaries, scoped credentials and a deployment shape that keeps a customer's data inside their own perimeter.

Observability & audit

Every agent action logged with its inputs, its tool calls and its result, so an operator can answer what happened and why.

Release pipeline

Model versions promoted only after the evaluation suite passes, with a route back to the previous version at any time.

How it has run

Five years committed, one launch date.

A research partnership cannot be run to a delivery plan, so it is run to evaluation gates instead. The dates below are the points where the programme changed shape.

January 2025

Partnership signed

The agreement that started the programme and settled the hard part up front: who owns the model, who owns the platform, and how a model version is judged fit to ship.
2025, first half

Benchmark before model

Before any training, the evaluation suite: a set of real agent tasks with a fixed way of scoring them, so progress could be argued from numbers rather than from demos.
2025, second half

First OK Model runs

Early training runs aimed squarely at tool calling and multi-step planning, measured against the suite and against the bilingual cases our market needs.
2026

Runtime and hardening

The agent runtime, the connector layer and the isolation model built around the improving model, with safety and prompt-injection work running alongside.
1 January 2027

Agentic OS launch

The CodeFix Agentic OS goes to market powered by the OK Model, with Philips Indonesia as its first client.
2027 – 2030

Committed research

The partnership is signed through 2030: successive model generations, a widening evaluation suite, and the research that keeps the product in front of what enterprises ask of it.

How it is built

Built so a customer can run it themselves.

The Agentic OS is sold to enterprises on security, privacy and controlled AI operations. That promise decides the architecture: the model and the runtime have to work inside a customer's own boundary, not only in ours.

Model

OK Model
PyTorch
Fine-tuning
Quantisation
Evaluation harness

Serving

GPU inference
Batching
Streaming
Model registry

Runtime

Agent scheduler
Tool registry
Task queues
State store

Controls

Tenant isolation
Scoped credentials
Full audit log
On-premise option

The benchmark decides

No model version reaches the product because it looked good in a demo. It ships when the evaluation suite says it is better, on tasks agreed in advance.

The data stays put

The architecture assumes a customer who will not let their work leave their own infrastructure, because in this market that customer is the normal case.

An agent is accountable

Autonomy without a record is not something an enterprise can buy. Every action an agent takes is attributable, inspectable and reversible by an operator.

What it gives us

A product that does not depend on anyone else's model.

A model we control
The Agentic OS launches on the OK Model, which we and AI Lab Bangladesh own. No third party can reprice it, deprecate it or change its behaviour under us.
The privacy claim is real
Because we hold the model, the product can run inside a customer's perimeter. The security promise is an architecture, not a policy document.
Progress that can be argued
The evaluation suite came before the model, so every claim about the OK Model is a number on a task list rather than an impression from a demo.
Two languages, one model
Bengali and English are handled by the model itself, which is what our market actually needs and what general models are weakest at.
Each side does its own work
Research and platform engineering are held by the teams that are good at them, with the benchmark as the contract between the two.
A horizon past the launch
Signed through 2030, the programme is funded to keep improving the model for three years after the product reaches its first customers.

Where it ships

The OK Model powers the CodeFix Agentic OS.

The Agentic OS is an enterprise operating system where you command many AI agents at once. They execute the tasks, coordinate the workflow and deliver the results, so your team spends its time on decisions instead of execution. It launches on 1 January 2027, powered by the OK Model, with Philips Indonesia as its first client.

See the Agentic OS

Security and privacy at the core Controlled AI operations, built for enterprise environments from the first line.
Powered by the OK Model Our own model, built with AI Lab Bangladesh, rather than someone else's API.
First client: Philips Indonesia The launch customer for the platform when it goes to market on 1 January 2027.

Want the Agentic OS in your business? We are taking early access requests ahead of the 1 January 2027 launch. Tell us what your teams spend their days executing, and we will show you what an agent should be doing instead.

Talk to us about it +880 1327-422480