[ Services ]
TERNARY designs and builds AI solutions, starting from the task, the data and the environment they will run in. Before any technology is chosen, we establish what the solution has to do, which data it works with and how the result will be judged.
Where needed, we cover the whole technical chain: data, model, runtime and hardware, software, integrations and deployment.
[ One solution, several layers ]
A typical project begins by defining the task and ends with a solution that can be used in real work. Every stage can be measured on its own and corrected where needed.
- We establish the task, the existing workflow, the constraints and the metrics by which the result will be judged.01
Scoping
- We assess the existing data: its quality, coverage, origin and the rights to use it. Where needed, we prepare a training or evaluation corpus.02
Data
- We decide whether an existing model is enough for the task, whether it has to be adapted, or whether training a specialised model is justified. The requirements of local, hybrid or cloud operation are taken into account as well.03
Choice of solution
- We adapt or train the model and evaluate it against a fixed test set and the metrics that matter for the task. Where needed, we quantise and optimise.04
Model and evaluation
- We determine the compute, memory and runtime environment required. The solution may run on a local workstation, on the organisation's server, as a hybrid, or in the cloud.05
Runtime and hardware
- We connect the model to an API, an application or an existing information system, and put the solution into the actual workflow.06
Software and integration
Before a project starts, none of this has to be known:
- which model to use
- whether the model needs adapting
- whether a separate model is needed
- how much GPU or memory is required
- whether to use a local server, a hybrid or the cloud
- how to connect the model to an existing system
These choices are made during the project, according to measured needs.
[Local where possible.Hybrid where useful.Cloud where justified.]
Local
Fits when the data has to stay under the organisation's own control, the load is continuous, latency matters, or dependence on an external service should be reduced.
Local AI does not necessarily mean a large server. Depending on the model and the load, the solution may run on a workstation, on a compact AI server, or as part of the existing infrastructure.
Hybrid
Fits when part of the work can be done locally, but some tasks need a larger model, additional compute or an external service.
Sensitive data and the steady load can stay local, and the cloud is used only where it brings a technical or economic benefit.
Cloud
Fits when the use is irregular, building one's own infrastructure is not justified, or the required model and compute are more sensibly used as a service.
The choice follows the task, the security requirements, the load and the total cost.
[ Where to use AI? ]
AI is practical above all where a larger volume of text, documents, audio or unstructured data has to be processed, where information has to be found, where a repeating workflow can be automated, or where existing software can be supported by a model-based function.
Depending on the task, an existing model may be enough, adapting one may be necessary, or training a smaller specialised model may be justified.
[ What we do ]
- 01
Models and data
We select, adapt and evaluate existing models, and where needed train smaller specialised ones. The work also covers preparing training and evaluation data, building corpora, checking their quality and documenting the rights to use the data.
- 02
Runtime and hardware
We size the compute and the runtime environment a model needs. The solution may run on a local workstation, on the organisation's server, in a hybrid setup or in the cloud, depending on the model, the load, latency and how the data has to be handled.
- 03
Software and integrations
We build APIs and web, mobile and desktop applications through which the model is connected to existing information systems, documents, databases and workflows. The model is one component of the solution, not the end result on its own.
[ Our own models and technology ]
Alongside commissioned work, TERNARY develops its own models and technology. HÄÄL is a project for Estonian text-to-speech (TTS), KUULA its speech-to-text (STT) continuation, and MATRIX the technical direction for compact models that can be run locally.
Developing them ourselves lets us judge training, data, quantisation, runtime and hardware from practical work, and apply the same experience when building other solutions.
[ Let us talk about the task ]
If you have a process, a dataset or a technical task where AI could be useful, the technical solution does not have to be known before getting in touch.
Describe the task and the current situation. We will help assess which model, runtime environment and software are justified for solving it.
[ Get in touch ]