Oragon Labs · Research
The lab behind the products.
Oragon Labs is our research arm. We work on one hard problem: making capable AI small enough, cheap enough and local enough to run where it is actually needed. We publish what we find, including the experiments that fail.
The research question
The best AI research in the world is useless to most of the world if it only runs in a data centre they can't reach.
So ours is not just can the model do it. It's can it do it here: on this hardware, in this language, on this budget, with the power off.
The lab
We are a research lab first. The products come out of the research, not the other way round.
Most of what the frontier produces assumes a data centre, a fibre line and a budget in dollars. Strip those away and most of the answers stop working. The lab exists to find the ones that don't: smaller models, inference that runs beside the data, languages the field has no commercial reason to serve, systems that survive a power cut.
A lab that only publishes its wins is a marketing department. We publish the failures too. They are usually the more useful half.
What the lab publishes
Field notes
Short, concrete write-ups of what we tried and what happened. A model on hardware that shouldn't run it, a language nobody has benchmarked, a deployment that broke in an interesting way.
Benchmarks
Numbers measured on the hardware institutions here actually own, not on an A100 nobody in the room has. Reproducible, with the setup published alongside.
Open weights
Models we train for low-resource languages and narrow tasks, released so the people they serve are not renting access to their own language.
Reference architectures
The deployment patterns that survived contact with a real hospital, bank or ministry, documented so the next institution doesn't start from zero.
[Field notes and benchmarks are not published yet. This is where they will live. See §9 of the positioning brief.]
What kind of lab
Plenty of labs research what AI could do. We research what it can do for people who are not being served.
That is a narrower question and a harder one. It rules out answers that need a hyperscaler, a fibre line or a Fortune 500 budget, which is most of the answers the field currently produces. What's left is the work we do.
Research
We start from a question a real institution has, not a benchmark. Most of our work is on efficiency: making capable models small enough, cheap enough and local enough to be deployable here.
Prototype
We build the smallest working system that proves or kills the idea, on the hardware the institution actually owns, in the languages its users actually speak.
Deploy
What survives becomes a system somebody depends on, installed close to the people it serves, owned by them, and supported by engineers in the same time zone.
Publish
We write up what we learned, including what failed. A lab that only publishes its wins is a marketing department.
Where it applies
One research agenda. Every sector that matters.
The same underlying work (small models, on-device inference, local languages, offline operation) unlocks a different problem in each of these. That's the leverage of doing it at the research layer.
Healthcare
Diagnostic and triage support that runs inside the hospital, offline, on hardware it already owns.
Education
Tutoring and teacher support in local languages, on the low-cost devices schools can actually afford.
Agriculture
Crop, pest and yield intelligence for smallholders, usable on a basic phone with no signal in the field.
Financial services
Fraud detection and credit intelligence tuned to local patterns, compliant with data-residency law by construction.
Public services
Citizen-facing services in the languages people speak, hosted on national infrastructure.
Accessibility
Speech, vision and communication assistance that widens access for people with disabilities, on affordable devices.
Safety & security
Video understanding that detects events on-site without streaming people's lives to a distant cloud.
Energy & utilities
Demand forecasting and fault detection for grids that are intermittent by nature, not by exception.
Logistics
Routing and supply-chain intelligence built for real road networks, informal markets and patchy connectivity.
The research agenda
What we're actually working on.
Four threads, chosen because each one removes a specific reason AI currently fails to reach people.
- Model efficiency
- Making capable models small enough to run on hardware institutions already own.
- On-device inference
- Moving computation to the data instead of the data to the computation.
- Low-resource languages
- Serious performance in languages the frontier labs have no incentive to serve.
- Offline-first systems
- Treating intermittent power and patchy networks as design inputs, not edge cases.
How we build
Sovereign. Close. Open. Within reach.
Four commitments, and each one is an engineering decision before it is a value statement.
Sovereign
Your data, your models, your jurisdiction, your control.
Close
Infrastructure and teams near the problems they solve.
Open
Built on open models, free of lock-in, resilient by design.
Within reach
Accessible and affordable in real-world conditions.
Put the research to work.
If you run an institution with a problem worth solving, or you want to build this with us, we'd like to talk.
Get in touch: hello@oragonlabs.com