Models are improving fast and the next bottleneck in AI is not intelligence but context. The winners will be whoever gives a model what a company actually knows, correctly and efficiently.

Vision

We think LLMs have created an opportunity to revisit neuro-symbolic AI. Researchers have explored combining neural networks with symbolic reasoning for decades, but the difficulty of building scalable systems, combined with the success of deep learning, kept it largely outside the mainstream.

For the first time, we may be able to build and maintain structured representations of real-world knowledge from unstructured information at a practical scale. That's the engine we are building. On Onyx's EnterpriseRAG-Bench of 21,000 documents, our system retrieved the required evidence for 96% of questions, compared to 44% for conventional RAG with a clear path to 100%.

The core of Jar is proven exact. Every answer carries the evidence it came from, and the same question on the same data always returns the same answer. This lets us guarantee correctness on retrieval and avoid hallucinations within a closed system. Underneath, Jar has achieved ACID transactions, which let people, agents and programs read and write the same model at once without ever seeing a half-updated company, losing a fact, or getting two answers to one question.

Contents

Jar is working with a limited number of companies to make their business processes run better. It connects to a company's systems and devices and builds the model of the business from everything they hold. People, applications and agents query that model directly, every answer is traced to its source, and it all runs on the company's own infrastructure.

Manifest

We are building in Waterloo and supported by Velocity and MaRS. We are working with a limited number of partners to build this with real company data. If you want to be one of them, write to founders@usejar.ai.