What I am actually trying to do here.
The lab's AI thesis is simple. The hard part of building useful AI products is not the model. It is the system around the model. Latency budgets, evaluation, memory, fallbacks, the human-in-the-loop UX. The experiments in this folder are aimed at those, not at the model itself.
The reason is leverage. The model is improving on someone else's roadmap. The system around it is the part the lab owns. Getting good at the second is more durable than betting on the first.
Model orchestration.
The orchestration layer routes a request to the right model based on cost, latency, and task. A short rewrite request goes to a small local model running on a laptop. A long-context reasoning request goes to a hosted frontier model. The router is the lab's, not the provider's.
Building this once means every AI surface the lab ships afterward can lean on the same primitive. That compounds.
- Local models for latency-bound and privacy-bound work
- Hosted models for reasoning-heavy work
- A router that picks based on task signature, not author preference
- A consistent evaluation harness across both surfaces
Local-first inference.
Running small models on consumer hardware is the experiment that has the most upside for the lab. Latency becomes a non-issue. Privacy becomes a non-issue. The economics flip from per-token to per-machine.
The experiments here cover quantization tradeoffs, the kind of tasks small models are actually good at, and the orchestration patterns that make them feel as capable as hosted models when paired with the right context.
A small model that ships beats a large model that is too expensive to call.
Agentic patterns.
The interesting agentic work is not 'the model can use tools'. The interesting work is 'the system around the model knows what to ask, in what order, with what budget'. The experiments here cover task decomposition, retry strategies, and the boundary between letting the model decide and letting the system decide.
Nyxera will inherit the patterns that prove themselves here. The rest will be left as lessons.
Evaluation, the part nobody talks about.
If you ship AI without a clear evaluation harness, you are guessing. The lab's experiments build small, project-specific eval sets that test the surfaces that matter. Not benchmark scores. Real workflows.
The discipline is to never ship an AI feature without an eval that proves it is doing what it claims to do. The discipline is also to keep the eval small enough that it actually gets run.