Product in development

Local AI, kept close.

We are defining a product for Norwegian law and accounting firms that need useful AI workflows and direct control over sensitive client data.

Join a discovery conversation →Contact Hakuya

Technology

Concept defined

Discovery

0 of 15 interviews

Product

Not built

Initial market

Norwegian professional firms with 2 to 50 professionals.

The first target segment is law and accounting firms. The buyer hypothesis is a managing partner or managing director, with the firm’s external IT provider as a likely technical stakeholder. Discovery starts in Trondheim.

Four planned workflows

Secure dictation and transcription

Turn meetings, calls, and dictated notes into reviewable text inside a customer-controlled environment.

Document drafting

Create drafts from the firm’s approved templates and source material, with a professional reviewing every output.

Case and client archive search

Retrieve relevant material across internal document collections without making a cloud service the system of record.

Accounting document extraction

Structure information from incoming records for review before it enters the firm’s accounting workflow.

Planned differentiation

One local package

The planned offer combines inference hardware, workflow software, and operational controls.

Professional-firm workflows

The product is planned around concrete drafting, search, transcription, and extraction work rather than a generic chat interface.

Norwegian context

Language quality, confidentiality expectations, and the working practices of smaller Norwegian firms will shape the acceptance tests.

Built for smaller firms

The goal is a repeatable offer for firms below the enterprise procurement tier.

Why local

Control is the product requirement.

The case for local deployment is durable: confidentiality duties, client expectations, direct data control, and lower exposure to changing cross-border and vendor risk. Cloud AI is not categorically illegal for the target firms. The thesis is that some workflows and client matters need a more controlled option.

NVIDIA technology plan

On-premises compute

NVIDIA RTX-class GPUs selected after the validated workflow, model, concurrency, and customer environment are known.

Language-model inference

CUDA as the accelerated computing platform and TensorRT-LLM for optimized LLM inference.

Deployment standardization

Evaluate NVIDIA NIM for packaged, consistent local model serving where the workload and licensing fit.

Speech workflow

Evaluate NVIDIA Riva ASR for local transcription, including Norwegian accuracy and domain-language testing.

This is a plan, not a statement of current deployment. No local or NVIDIA inference is running today.

Evidence before build

Fifteen conversations decide what happens next.

Continue only if at least 3 of 15 interviews qualify.

A qualifying interview identifies a budget owner, a recurring costly workflow, and either a cloud blocker or willingness to pay for a local alternative.

The product has not been built. There are no customers, pilots, prototypes, or deployed product models to claim. The first pilot begins only after the discovery gate, with one customer and one proven workflow.

Discovery

Tell us how your firm handles sensitive work →