Technology
BioTekk combines conversational AI, longitudinal modelling, configurable safety processes and secure service architecture. The system is being developed as a set of controlled modules so that capabilities can be evaluated, governed and deployed according to their intended purpose.
Controlled modules
Separating capability into layers is a governance decision as much as an engineering one. It allows a function to be tested against its own acceptance criteria, restricted to the deployments where it is appropriate, and updated without silently changing the behaviour of everything around it.
The conversational layer is intended to understand everyday language, retain relevant context within authorised boundaries and generate responses that are supportive, proportionate and transparent.
It is not designed to imitate a named clinician or conceal that the user is interacting with an AI system.
BioTekk’s modelling approach is intended to examine multiple forms of information, including self-report, language patterns, interaction history and, where separately authorised, contextual or biometric data.
The purpose is to identify change and support further enquiry. Model outputs represent probabilistic estimates and should include uncertainty.
Repeated observations may provide more useful information than a single interaction. The platform is therefore designed to evaluate trajectories, variability and changes from an individual baseline.
This approach requires careful controls to avoid treating ordinary variation as pathology.
A dedicated safety layer is intended to evaluate interactions against clinically reviewed rules and model-assisted indicators.
The architecture should support different response levels, including immediate safety messaging, encouragement to seek urgent help, notification to an authorised team where permitted, and documented human review.
Where the system provides psychoeducation or service information, approved knowledge sources can be maintained separately from the generative model.
This supports source control, updating, auditability and reduction of unsupported output.
BioTekk is being developed using separable services for identity, consent, conversation, analysis, safety, notifications, reporting and integration. Modularisation supports controlled testing, restricted access and independent updating.
Specific infrastructure and hosting claims should be published only after the production design and suppliers are confirmed.
Service boundaries
Integration approach
The intended integration model is API-led and standards-aware. Potential healthcare integrations may include structured exchange using relevant interoperability standards, so that information can move between systems in a form each side can interpret reliably.
Integration is not only a technical exercise. Each interface requires an agreed data flow, a lawful basis, an information-governance review, defined error handling and a support model for when something goes wrong in live service.
Compatibility with a named electronic patient record should not be claimed until the interface has been implemented, tested and approved with the relevant supplier and deploying organisation. Where this site describes integration, it describes intent and method — not completed certification.
Model evaluation
Evaluation should include, at minimum, the dimensions below. Performance should be reported for a specified model version, dataset, population and intended use — a figure without that context is not evidence.
Performance for each defined safety task, reported separately rather than as a single aggregate accuracy figure.
Analysis of both error directions, including the operational consequence of each and the capacity required to absorb it.
Differential performance across relevant subgroups, identified by active evaluation rather than assumed absent.
Structured review of whether responses are proportionate, bounded and consistent with clinical guidance.
Testing for unsupported or fabricated output, particularly in psychoeducation and service-information responses.
Behaviour under ambiguous, adversarial, distressing or out-of-scope input, and under unusual interaction patterns.
Response latency and platform availability measured against defined service-level objectives.
Assessment against recognised accessibility criteria, including assistive-technology testing.
How real users and professionals interpret and act on outputs, including foreseeable misuse and alert fatigue.
Every performance figure BioTekk publishes is expected to carry its model version, dataset description, population, intended use and evaluation date. Targets currently shown on this site are internal engineering objectives, not validated clinical outcomes.
Human oversight
Human oversight is required at product, model and service levels. This includes review of safety policies, approval of material model changes, monitoring of incidents and near misses, investigation of performance drift and the ability to suspend or modify automated functions.
Technical, clinical and information-governance teams are welcome to interrogate the design. We would rather answer difficult questions early than defend vague claims later.