Applied research lab

Computing systems for the AI era

Today's computing systems will undergo a massive transformation to efficiently and scalably serve AI workloads. NiNI Labs is an applied research lab dedicated to envisioning the next generation of these systems.

Why the stack has to change

Three shifts that make general-purpose infrastructure the wrong foundation for AI.

Inference is the workload

For the last decade the hard problem was training one model. Now the hard problem is serving millions of model calls a day, each with a different shape, latency budget, and cost ceiling. The systems underneath were never designed for that.

Agents change the shape of compute

An agent is not one request. It is a long-running loop of model calls, tool invocations, searches, and state. Bursty, branching, and stateful — the opposite of the stateless request/response that today's serverless platforms assume.

Hardware outran the software

Accelerators sit idle between batches, memory bandwidth is the bottleneck, and schedulers still treat a GPU like a CPU. Closing that gap is a systems problem as much as a hardware one.

Research areas

Where the lab spends its time. Each area is measured against live workloads from the Nicrron gateway and the Ninoo agents.

Serverless agent runtimes

Execution models where an agent's stages — model calls, tool use, search, custom code — are scheduled independently, scaled to zero when idle, and resumed without losing state.

Inference scheduling and orchestration

Placing requests across heterogeneous accelerators, batching without hurting tail latency, and routing between models by cost, quality, and load in real time.

Memory and state for long-lived agents

How an agent remembers across hours or weeks: durable conversation state, retrieval that stays fresh, and caches that cut repeated context without changing answers.

Compound AI systems

Pipelines that chain multiple models with non-model stages such as search, ranking, and structured data — and the tooling to import, observe, and optimize them as one unit.

Efficiency per token

Energy, memory, and dollars per useful token. Speculative decoding, quantization, KV-cache reuse, and co-scheduling, measured end to end rather than on a benchmark harness.

Evaluation and observability

Instrumentation for multi-agent systems: tracing every stage, attributing cost, catching regressions, and turning production traffic into the next experiment.

Projects & Services

Research that holds up becomes a service. This is the first.

NiNI Cloud

Waitlist

Serverless inference for AI agents. Run everything from simple agents to complex multi-agent systems with custom logic and data sources. Import existing agentic pipelines, chain multiple models with non-model stages such as search and custom data sources, and scale it all seamlessly. The platform handles scheduling, orchestration, and optimization, so you can focus on adding new capabilities to your agents.

  • Any agent shape: single loop, multi-agent, or a pipeline with custom code
  • Import what you already run — no rewrite to get started
  • Model and non-model stages in one graph: search, retrieval, your data
  • Scheduling, orchestration, and optimization handled by the platform

How we work

1

Build in production

Our test bed is real traffic. The Nicrron gateway and the Ninoo agents give the lab live workloads to measure against, so ideas are judged on tail latency and cost, not on slides.

2

Systems over models

We do not train frontier models. We work on everything around them — the runtime, the scheduler, the data path, the metal — where most of the remaining efficiency lives.

3

Ship what works

Research that holds up graduates into a product. NiNI Cloud is the first: the serverless agent runtime that came out of the lab's scheduling and orchestration work.

One lab, one stack

The lab sits between the products that generate the workloads and the facilities that run them.

The proving ground

Nicrron

The unified gateway routes production traffic to 65+ models. It is where the lab's routing, caching, and failover ideas get tested first.

nicrron.ai
The first agent workload

Ninoo

A coding agent in the terminal and a browser builder — long-running, tool-heavy, stateful. Exactly the kind of workload NiNI Cloud is being built to run.

ninoo.ai
The physical layer

Nicroo Core

AI data centers built for AI. The lab's findings on accelerator utilization and scheduling feed directly into how Nicroo Core facilities are designed.

Nicroo Core

Common questions

Is NiNI Labs a company or a research group?

An applied research lab inside the Nicrron family. It operates like a lab — hypotheses, experiments, measurements — but its output is products and infrastructure, not just papers.

What has the lab shipped?

NiNI Cloud, serverless inference for AI agents, is the lab's first service and is currently taking waitlist signups. Routing and caching work from the lab already runs inside the Nicrron gateway.

Do you publish?

We intend to share findings as write-ups and open tooling where it does not compromise customer workloads. If you are working on similar problems, we would like to hear from you.

Can I collaborate or join?

Yes. Email support@nicrron.ai with the subject "NiNI Labs" and tell us what you are working on — research collaborations, design partnerships for NiNI Cloud, and engineering roles are all open conversations.

Working on the same problems?

Research collaborations, design partners for NiNI Cloud, and engineers who want to work on the runtime — all welcome.

Get in touch