mission
agni.works builds and studies small, specialized models that run where the data lives — on owned hardware, under the operator's control, against problems defined precisely enough that scale stops being the deciding variable.
The premise: model capability and model utility are not the same thing. For a large class of real problems — perception on a vehicle, analysis of proprietary footage, scoring against an expert's rubric — the deciding factors are ownership, privacy, latency, and cost. A frontier model accessed through an API fails those constraints before the first token is generated.
The competitive thesis
A small lab cannot outcompete frontier labs at general capability, and should not try. Two paths remain viable, and agni.works works both:
- Find the niche — domains where general training data is thin and domain signal is everything: biomechanics, driver behavior, affect over time. A specialized model with the right conditioning signal competes on ground the generalist never trained on.
- Reframe the product — instead of a bigger model answering every question, a small model owned by the client, tuned to their task, running on their hardware, seeing their data and no one else's. The product is not intelligence-in-general; it is a specific competence, delivered with control.
The research bet underneath both paths — the one REVEAL exists to test — is that exogenous, domain-specific signals can close enough of the capability gap that a small conditioned model stands against generalist giants on the tasks that actually matter to an operator.
Constrained hardware as a discipline
Everything here runs on consumer and edge hardware — dual consumer GPUs for training, Jetson-class devices for deployment. That is partly economics, but mostly method: constraint forces architectural honesty. When the memory budget is fixed and the interconnect is asymmetric, every design decision has to be justified against the signal it preserves, not the convenience it buys.
The practical consequences are documented across this site: training topologies chosen for what they make observable rather than what they make fast, inference pipelines budgeted frame by frame, and models sized to what a client can actually own and operate.
Privacy and control
Data that never leaves the device. Models the operator owns outright. This is a personal operating principle before it is a market position — but it is also a market position: for clients whose footage, subjects, or methods are themselves the competitive asset, a locally-run specialized model is not a compromise against the cloud frontier. It is the only acceptable architecture.
Some of what runs here will stay a first attempt; some of it will become product. Either way, the mindset is the same — approach the purpose differently from the start, and let the giants keep the general case.
Contact
For collaboration or correspondence: Arjun.Joshi@Agni.works