Solutions
Small models, packaged as solutions
One idea in four parts: find the narrow task, distill a compact model that holds it, package it to embed and install where the work already happens, and prove it against a baseline. Each pillar below sets out what that means and how an engagement actually runs.
Purpose-built, not general-purpose
Domain AI Models
The work of the company is building AI models for narrow, high-value tasks inside regulated domains: classifying transactions, extracting fields from statements and documents, matching and reconciling records, triaging intake. Each model is trained for one niche use case, on vetted domain data, and is measured against a known-good baseline before anything depends on it.
None of it is generative AI. There is no chatbot, no content generation, and no frontier-model ambition anywhere in the roadmap. A task model either performs its task at the reliability bar the domain demands, or it does not ship — and getting compact models to that bar is precisely the research program.
What that covers
- Transaction and record classification
- Field extraction from statements, invoices and forms
- Record matching and reconciliation support
- Domain-specific document triage
How the models get small
Teacher–Student Distillation
The method is distillation. Capable teacher models are used under controlled conditions during training to generate supervision for a much smaller student model dedicated to a single task. The student learns the task, not the teacher’s generality — which is exactly the trade a regulated deployment wants.
What leaves the lab is the student: compact, evaluated against a fixed test suite, with no dependency on any external AI service at run time. That independence is what makes on-premises and air-gapped deployment realistic engineering rather than a marketing line.
What that covers
- Task-specific student model training
- Curated and synthetic training-data construction
- Fixed evaluation suites against known-good baselines
- Models sized for ordinary on-premises hardware
A solution, not a science project
Embeddable & Installable
A model is only useful once it is inside a workflow. Every model is packaged as a component with a documented interface, so it can be embedded into an existing platform — a bookkeeping system, a claims pipeline, an intake tool — rather than asking anyone to adopt a new one.
Deployment profiles run from locally installed but connected, through to fully air-gapped: no external calls at all, with data ingested in batch inside the boundary. The strictest profile is a design input from day one, not a retrofit — and the difference between the profiles is stated plainly, never blurred.
What that covers
- Packaged models with documented integration interfaces
- On-premises installation on client hardware
- Fully air-gapped deployment profile
- No run-time dependency on external AI services
Where an engagement starts today
Use-Case Discovery & Pilots
The models are in development; the discovery is deliverable now. We work with organizations in accounting, banking, finance, taxation, healthcare and legal to identify the specific tasks that fit this technology: narrow in scope, high in volume, checkable against ground truth, and constrained by data that cannot leave the building.
A candidate use case that survives scoping becomes a pilot with a defined baseline and a defined bar. One that does not survive scoping gets a plain "this is not a fit" — which costs a conversation, not an engagement.
What that covers
- Niche use-case identification and scoping
- Data-sensitivity and deployment-profile assessment
- Pilot design with measurable baselines
- Integration path into existing platforms
Have a task you think a model could hold?
Every engagement starts with use-case discovery. Describe the workflow in plain terms and we will tell you whether it fits what compact task models can do — or that it does not need us at all.