Inspiration

AI has made remarkable progress in reasoning, coding, and conversation. However, completing real-world work still requires coordinating multiple tools, managing context, verifying results, and handling failures.

We created Aria to explore a different interaction model: an AI workspace where users describe an objective and the system helps turn that intent into a structured, verifiable workflow.

What it does

Aria is an agentic AI workspace developed by Arthera. It brings together three complementary experiences:

Aria Chat: Understand requests, explore information, and develop plans through natural conversation.
Aria Work: Coordinate multi-step tasks, analyse information, invoke tools, and produce actionable deliverables.
Aria Code: Inspect repositories, generate and modify code, assist with testing, and support developer workflows.

Aria also integrates specialised financial intelligence, including quantitative research, portfolio analysis, risk modelling, and backtesting.

For this hackathon, our focus is a complete agentic workflow: understand a user objective, plan the required steps, invoke specialised tools, inspect outputs, and deliver a result the user can review.

How we built it

Aria uses a modular architecture connecting AI models, orchestration services, domain-specific engines, and user-facing applications.

Our existing platform includes streaming model responses, tool calling, multi-step agent loops, hybrid cloud/local inference, and specialised quantitative analysis.

The hackathon implementation focuses on integrating NVIDIA open models through Nebius Token Factory into Aria's agent orchestration layer. The goal is to use open-model inference for reasoning and task planning while retaining independent control over execution, tool permissions, and result verification.

This approach keeps the application model-flexible and separates inference from real-world actions.

Challenges we ran into

Building reliable agents involves more than generating high-quality responses. Important engineering challenges include maintaining task state across multiple tool calls, recovering from failed operations, controlling execution permissions, and preventing generated outputs from being treated as verified results.

Another challenge is balancing reasoning quality, latency, and inference cost across tasks with different levels of complexity.

Accomplishments that we're proud of

Aria has evolved beyond its initial financial AI foundations into a broader agentic workspace.

Its technical foundation combines an extensible orchestration architecture, developer tooling, structured tool execution, and a quantitative engine with portfolio optimisation, risk modelling, and historical backtesting capabilities.

The key accomplishment is connecting general-purpose AI capabilities with specialised computational tools that can produce inspectable outputs.

What we learned

The most important lesson is that agent reliability depends on the entire system, not only the underlying language model.

Planning, tool selection, execution isolation, validation, state management, and user oversight are all essential components of trustworthy autonomous workflows.

What's next for Aria

Our roadmap focuses on longer-running agent tasks, more robust execution environments, persistent project context, model-efficient inference, and stronger evaluation systems.

Longer term, we aim to extend the same principles from digital workflows toward multimodal systems and physical AI, where reliable perception, planning, and action become increasingly interconnected.