Inspiration

AI is advancing rapidly, but users often face a difficult choice: run models locally for privacy and control, or rely on cloud-based AI for greater accessibility and computational capabilities.

We wanted to eliminate this trade-off. Our inspiration was to build a unified AI workbench that combines the privacy and independence of local models with the intelligence and scalability of cloud models. This led to Knovia AI, designed to make AI more flexible, accessible, and adaptable to different workloads.

What it does

Knovia AI is an Adaptive Hybrid Intelligence Workbench that supports two operating modes:

Local Mode: Uses Ollama and locally hosted models for private, controlled AI execution.
Cloud Mode: Uses NVIDIA Nemotron through Nebius Token Factory for cloud-based AI inference.

The platform brings multiple AI capabilities into one environment:

Conversational AI and general LLM interactions.
AI agents for task execution and orchestration.
Coding assistance through an AI coding agent.
Centralized model and provider management.
Configurable model routing based on deployment mode.
Transparent visibility into the models and providers handling requests.

Its architecture is designed to let the same application operate locally or in the cloud without maintaining separate versions.

How we built it

We developed Knovia AI using a modular architecture that separates the application, orchestration, and model execution layers.

We retained Ollama as the local inference provider and extended the provider architecture to integrate NVIDIA Nemotron through Nebius Token Factory using its API.

A centralized configuration determines the operating mode. In local mode, AI requests are handled by Ollama. In cloud mode, requests are routed through Nebius to NVIDIA models.

We also focused on integrating the provider selection across conversational AI, general LLM operations, agents, and the coding assistant, so that all AI features follow the same deployment configuration.

Challenges we ran into

One of our biggest challenges was extending an existing local AI architecture to support cloud inference without breaking its existing functionality.

We had to address differences between local model execution and cloud API-based inference, provider configuration, environment variables, API communication, and deployment requirements.

Another challenge was ensuring that every AI component—including chat, agents, and coding assistance—uses the correct model provider rather than accidentally falling back to a local model in cloud mode.

We approached these challenges by focusing on centralized provider management and consistent model integration.

Accomplishments that we're proud of

We are proud of building a unified AI workbench that brings local and cloud AI capabilities together rather than treating them as separate applications.

Our key accomplishments include:

Preserving local AI capabilities through Ollama.
Integrating NVIDIA Nemotron through Nebius Token Factory.
Designing a shared provider architecture for local and cloud execution.
Bringing conversational AI, agents, and coding assistance into one platform.
Making deployment mode configurable without requiring separate application codebases.

Most importantly, we created an architecture that can evolve as new models and inference providers become available.

What we learned

We learned that building a useful AI platform involves much more than connecting language models to a chat interface.

Model abstraction, provider management, inference reliability, deployment architecture, and consistent orchestration are essential to building a flexible AI system.

We also learned how to integrate cloud inference with existing local infrastructure while maintaining a consistent experience across different AI capabilities.

What's next for Knovia AI — Adaptive Hybrid Intelligence Workbench

Our next goal is to make Knovia AI more intelligent in how it selects and coordinates models.

We plan to introduce:

Adaptive Model Routing: Select models based on task complexity, capability, latency, and inference cost.
Multi-Model Orchestration: Divide complex requests into subtasks and assign them to suitable models.
Quality-Gated Escalation: Escalate tasks to more capable models when initial results do not meet quality requirements.
Cost and Performance Benchmarking: Compare adaptive routing against fixed-model execution.
Expanded Agent Capabilities: Improve specialized agents and coding workflows.
Scalable Cloud Deployment: Make the workbench accessible to teams through secure, reliable cloud infrastructure.

Our long-term vision is to make Knovia AI a unified intelligence platform where users can benefit from local privacy and cloud intelligence while the system intelligently manages how AI resources are used.