Designs, builds and tests the workflow
- Designs the workflow with your team
- Connects your data and the models
- Sets who can send which data to which model, and when to switch models
- Tests results on your own cases
- Helps you run and monitor it
AgenTek Labs
Would a small model you own beat the API you pay for?
We test it on your own past cases before a production build. If it meets your quality, cost and data requirements, we build it into your workflow. If it does not, you have the evidence to keep your current approach.
Who does what
Models are the easy part. The workflow around them is the work.
Serves the models
- GPUs, scaling and uptime
- The model endpoints the workflow calls
- You choose the provider and hold the account and the bill
Owns the outcome
- The workflow and the data
- What counts as a good answer, and which failures are unacceptable
- The final decision to go live
Works with the inference provider you choose, a frontier API, or your own cloud. You hold the accounts and the bills.
Start with one workflow
One workflow, tested properly, before anything is scaled.
Describe the job and the answer you need
Agree what good looks like and which failures are unacceptable.
Compare the options
Your current frontier model, an open model as it is, and a tuned model only where the numbers justify it.
Build a limited integration
Connect the real data sources and the chosen models for that one workflow.
Report and decide
Quality, cost per 1,000 cases, speed and critical failures for each option, with a recommendation: scale it, change it or stop.
What you get: a written report, the tested set of cases, the working limited integration, and a signed passport for any model we recommend. Scope is agreed with you up front.
What we can build
Fine-tuning is one tool. The workflow needs the rest.
- Model comparison on your own cases
- Retrieval and context design
- Routing and fallback rules between models
- Fine-tuning where it pays
- Access and data rules
- Monitoring, cost tracking and change checks
- A signed passport for each model version
Control, in plain words
Four questions your risk team will ask. You will have the answers.
Who can send which data to which model?
Rules by role and by type of data. For example, customer records go only to models inside your cloud.
Where does each model run?
In your cloud, with a UK or EU provider on GPUs or specialised inference hardware, or through a frontier API. We check residency, contractual controls and model support for the chosen route.
What does it cost?
Cost per 1,000 cases for each model, measured in testing and tracked once it runs.
How do you know when a model fails or changes?
We re-run your locked test set whenever a model or its version changes, and each approved version carries a signed passport.
Routing tools pick a model for each request. We give you the tested rules and the evidence behind them, across the whole workflow.
In a pilot: your own sign-in, four roles, an audit trail, signed jobs, a locked test set, and containers that run in your cloud.
UK and EU inference advice
Find the right home for your AI workload.
Bring your model, expected usage and data rules. We help you compare hosting routes, then connect the chosen service to your workflow.
Define the requirements
Model and licence, data location, throughput, response time, security and support. UK or EU hosting is assessed alongside the contract and operational controls.
Compare viable providers
Check GPU or specialised inference hardware, support for your model weights, capacity and total operating cost. Availability and compatibility are confirmed with the provider.
Keep control of the service
You choose the provider and contract with them directly. You hold the account and the hosting bill; AgenTek scopes its advice and integration separately.
If a provider pays AgenTek a referral fee, we disclose it before you decide.
Who you will work with
Chandra leads your engagement.
AgenTek is a senior-led UK consultancy. Chandra brings architecture experience across AI, data platforms, cloud and identity, including work in regulated sectors.
He works with your team on the business case, evaluation, integration and handover. Specialist delivery support is agreed with you as part of the scope.
For inference and cloud providers
You provide the infrastructure. We help the customer put it to work.
We are looking for referral and co-sell relationships with providers serving UK and EU customers.
What AgenTek brings
- Client-side discovery and the business case
- Model evaluation on the customer's own cases
- Data, workflow integration and handover
- Support for UK teams with residency and governance requirements
Workloads to qualify together
- Task-specific open models, often in the 1B to 8B range
- Document extraction, triage and drafting
- Deployment on one GPU or specialised inference hardware, where compatible
- UK or EU hosting, subject to the customer's requirements
A clear working relationship
- The customer holds the provider account and bill
- We agree responsibilities before an introduction
- We seek deal registration for referred opportunities
- Joint activity and customer references are agreed case by case
How we test: ModelFit
Every option sits the same exam.
ModelFit is the tool we use to test each option on cases it has never seen. The test set is locked before any model is tried, and each approved model gets a signed passport your risk team can check.
Try the Studio demo
Runs in your browser with sample data. It shows the testing steps; it does not train models.
Own the capability, not the vendor's roadmap
Plan your model migration before the deadline.
OpenAI ends new self-serve fine-tuning jobs for active existing customers on 6 January 2027. Listed GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini snapshots, including their fine-tunes, retire on 14 January 2027. Check the affected models and dates. We can test a replacement API or an open model against your own cases before you switch.