The company behind Annie
Annie is built by Evari, a fintech company with over a decade of experience building technology for complex, regulated industries. We understand the gap between impressive technology and real operational reliability, because we've spent ten years navigating it.
That experience taught us something important: in regulated environments, the cost of a wrong answer isn't a bad user experience; it's a compliance breach, a financial loss, or a decision that affects lives. When we turned our attention to AI in mid-2025, we brought that standard with us.
What we saw was wrong with frontier AI
The frontier labs built extraordinary systems: ChatGPT, Claude, Gemini. Impressive feats of engineering at planetary scale. We dug into them deeply. And the more we understood how they were built, the clearer the problems became.
- Context window limits Lose the thread on long documents, complex workflows, and multi-step decisions. Not a fixable feature; it's a fundamental architectural constraint.
- RAG accuracy problems Retrieval-augmented generation introduces retrieval errors that compound at enterprise volume. The model doesn't know what it doesn't know.
- Cost and power at scale Economically unsustainable for most organisations at any serious volume. The per-token economics only work if you're not actually using the system heavily.
- No state management Every session starts from zero. No memory of prior interactions, no accumulating understanding of your domain, no improvement over time.
The research community was already reaching a conclusion the frontier labs don't want to hear: small, domain-tuned specialist models, verified by independent peers and deployed on infrastructure you control, outperform general-purpose giants at a fraction of the cost. We decided to build that architecture.
Read our full thinking on why Annie is differentWhat started as an experiment in the attic
We didn't set out to build a platform. We set out to understand the problem properly, from the inside. That meant building our own infrastructure and training our own models, not wrapping someone else's API and calling it sovereign.
The first proof of concept was built in our own attic. Not in a commercial data centre — in an attic, using consumer RTX 5090 GPUs. Hardware that costs a fraction of enterprise-grade AI chips and draws minimal power by comparison. If the architecture was sound, it would work here first.
It did.
What made the Annie architecture possible wasn't just the models; it was the communication backbone. Evari had already developed proprietary data streaming technology as part of our fintech work, built to reduce reliance on high-cost cloud services and increase throughput and resilience. That infrastructure became the backbone of Annie's agentic framework: the layer that coordinates the classification, routing, judgment, and verification pipeline at the core of how Annie thinks.
The base Annie workforce models followed: a proprietary foundation model trained from scratch on sovereign data. Not fine-tuned from a frontier model. Not dependent on external weights. Trained, owned, and deployed entirely by us.
The proof of concept became a platform. Annie became real.
How we got here
Annie didn't start as the plan. Our original hypothesis was that the world needed a better way to build AI-powered operational workflows: enterprise-grade secure, but accessible to the business users who actually needed them. That became QuivaWorks: agentic orchestration for operational workflows, built for human-centric advisory businesses rather than AI engineers.
But as we built QuivaWorks, we found ourselves going deeper into the models underneath it. The more we worked with frontier models, re-engineering how they were implemented, testing their limits in real enterprise environments, the clearer it became that a better architecture was possible. More accurate for specific domains, more cost-efficient at scale, and more appropriate for the high-stakes work enterprises actually needed to do. That parallel R&D became Annie.
And then the broader picture came into focus. What started as an internal technical question has become a fundamental one for society: are we going to allow US-led hyperscalers to control our AI future? Or will the world build sovereign, compliant solutions that can compete, and outperform frontier models, for the specific purposes that matter most?
What began as in-house R&D is now a direct response to that question.
The Team
Annie was built by a team that combines deep enterprise domain expertise with the technical conviction to build AI from first principles, not just deploy it.
Chairman. Decades of enterprise leadership in regulated industries gives Daniel an uncompromising view of what AI must do to be trusted in high-stakes environments.
CEO. Leads operations and enterprise delivery, ensuring the platform development stays grounded in real-world deployment requirements and that Annie's commitments to partners and customers are kept.
CTO. Architected Annie's orchestration framework and oversaw development of the Workforce Foundation AI model from scratch, including building the proof-of-concept data centre in our own attic using consumer hardware. 20+ years designing distributed infrastructure, cloud compute, and data streaming systems at hyper scale
