GERA-1

GERA-1 Gerald is a multi-agent AI business operating system where specialist agents collaborate on research, finance and operations, keeping humans informed and in control.

Comment

Inspiration

GERA-1 Gerald began with a simple question: What if a small business could have its own team of AI specialists, working together rather than relying on one general-purpose assistant?

As the founder of A Moment of Escape Ltd, I wanted to build something that would help independent business owners manage the many different responsibilities involved in running a company.

My background in mathematics, education and computing also influenced the project. I wanted an AI system that supports exploration, understanding and decision-making rather than simply delivering answers.

The result is Gerald: a characterful, multi-agent business operating system designed around collaboration, transparency and human control.

What it does

GERA-1 Gerald brings together specialist AI agents within an interactive business interface.

Toby acts as the central orchestrator, coordinating requests and delegating tasks to the appropriate specialists.

The system currently demonstrates several capabilities:

  • AI conversation: Toby can communicate through NVIDIA Nemotron, accessed using Nebius Token Factory.
  • Business decision support: Bobathan and Saul collaborate on a synthetic brownie-production scenario, combining ingredient availability with financial checks.
  • Research: Percival uses the Tavily API to investigate topics and return sourced findings with clickable references.
  • Content creation: Montgomery receives Percival's research and prepares content drafts for human editorial review. The interface distinguishes AI-generated content from offline template previews.
  • Document export: Montgomery's workflow includes a locally generated PDF containing draft content and research sources.
  • Interactive interface: Gerald features an animated medallion, character-based navigation and specialist workspaces.

The system is designed to be transparent about what has actually happened, what is simulated and what still requires human approval.

How we built it

GERA-1 was developed primarily in Python and JavaScript, using a local backend connected to a custom browser-based interface.

The architecture separates orchestration, specialist agents, business workflows, provider adapters and access-control policies.

Key technologies include:

  • NVIDIA Nemotron-3-Ultra-550B-a55B, accessed through Nebius Token Factory, for live AI conversation and supported content-generation workflows.
  • Tavily API for live, source-based research.
  • Python for backend services, workflow coordination, validation and testing.
  • HTML, CSS and JavaScript for Gerald's interactive interface.
  • GitHub for source-code publication and version control.

The project was developed using an AI-assisted coding workflow, with iterative testing, debugging and human review.

Privacy and safety were considered throughout development. API credentials remain server-side, synthetic business data is used for demonstrations, and offline functionality is preserved where possible.

Challenges we ran into

One of the biggest challenges was connecting a visually ambitious interface to genuine backend functionality.

It was important that Gerald did more than display attractive animations. When an agent appeared to research, calculate or collaborate, I wanted the system to distinguish real operations from demonstrations.

Integrating multiple providers introduced further challenges, including intermittent network connectivity, unavailable API responses and differences between live AI output and offline fallback behaviour.

Another challenge was coordinating research and content generation without losing source attribution or allowing unsourced information to appear as verified research.

We also had to address interface issues, including content being clipped in the display, while preserving the existing design and animations.

Finally, working within the hackathon deadline meant prioritising a functioning, demonstrable system over implementing every planned feature.

Accomplishments that we're proud of

I'm particularly proud that Gerald has progressed from a concept into a functioning, publicly available open-source software project.

Highlights include:

  • Building an original, character-driven interface with an animated medallion and individual specialist identities.
  • Successfully connecting Toby to NVIDIA Nemotron through Nebius.
  • Demonstrating genuine Tavily research through Percival, including accessible source links.
  • Creating a research-to-content workflow involving Toby, Percival and Montgomery.
  • Implementing a synthetic business decision-support scenario involving multiple specialists.
  • Preserving clear distinctions between live AI responses, offline templates and simulated workflows.
  • Adding PDF export for editorial review.
  • Establishing an automated test suite, with 173 Python tests passing in the latest reported run.

Most importantly, Gerald demonstrates an approach to AI in which specialist agents work together while the business owner retains responsibility for decisions.

What we learned

This project reinforced that successful AI systems require more than powerful language models.

Orchestration, transparency, reliability and interface design are equally important.

I learned how to connect specialist workflows to external AI and research services, handle unavailable providers, preserve source attribution and design useful offline behaviour.

I also learned the importance of testing what users actually experience, not simply whether backend functions execute successfully.

Perhaps the most valuable lesson was that AI development can be a collaborative, iterative process. Human creativity and judgement remain central, even when AI tools contribute substantially to implementation.

What's next for GERA-1

Gerald is intended to grow into a practical operating system for small and independent businesses.

Planned developments include:

  • Expanding the specialist agents' capabilities and collaboration.
  • Connecting inventory, orders and finance to authorised real business data.
  • Developing richer, branded newsletters and marketing documents.
  • Improving PDF presentation and editorial workflows.
  • Adding more sophisticated research, provenance and audit capabilities.
  • Strengthening reliability, deployment and onboarding.
  • Exploring how Gerald could support other small businesses beyond A Moment of Escape Ltd.

The long-term vision is a business operating system that combines the usefulness of AI with the personality of a collaborative team.

Gerald isn't intended to replace the business owner. He's intended to give them a team.

Built With

Share this project:

Updates

Submission history