Ari

Ari is an AI work OS that turns everyday requests into verified CRM, task, email, calendar, and meeting actions so teams get outcomes, not just chat replies.

Comment

Inspiration

Teams still jump between CRM tools, task managers, email, calendar, meeting notes, and chat. We wanted to build one place where people can ask for work naturally and actually see the result happen in the product.

What it does

Ari is an AI work OS for modern teams. It helps with CRM, tasks, reminders, team workflows, email, calendar, meetings, files, and memory from one workspace.

A user can say, “Group these leads by segment and prepare a follow-up campaign,” and Ari turns that request into real actions. It does not just give a chat response; it updates the relevant workspace and shows what changed.

Ari also carries a safe summary of the current workflow across supported AI providers. This means users can switch between Gemini, Vertex, OpenRouter, or Codex without having to explain the whole task again.

How we built it

We built Ari with Next.js, Node.js, PostgreSQL, Electron, Python, and an Agno-based agent runtime.

The AI model plans the task, while our backend handles the actual execution. Every tool call is validated, scoped to the active user and session, checked against permissions, and recorded before Ari reports success.

Challenges we ran into

The hardest part was making the agent reliable with real work.

Natural prompts can be vague. Models can send invalid tool arguments, repeat an action after a timeout, or say a task is done when the product did not change. We solved this with typed tool contracts, session-aware context, confirmation gates, idempotent actions, run tracking, and verified tool outcomes.

We also had to make active sessions independent, so one long-running task does not block another. Users can steer an active task with a new instruction while keeping the original context.

Accomplishments that we're proud of

  • Built a full AI workspace instead of a single-purpose chatbot.
  • Connected natural-language workflows to CRM, tasks, teams, email, calendar, meetings, and files.
  • Made agent actions visible in the actual product.
  • Added safe context sharing across Gemini, Vertex, OpenRouter, and Codex.
  • Added session isolation, in-flight steering, confirmations, and duplicate-action protection.
  • Built a dashboard experience that makes AI work easier to review and trust.

What we learned

We learned that a useful agent needs more than a good model. It needs reliable tools, clear ownership of data, good error handling, confirmations, and a way to prove that the work actually happened.

We also learned that people rarely give perfect prompts. The product has to understand context, handle follow-up instructions, and ask for clarification when needed.

What's next for Ari

Next, we want to improve more end-to-end workflows, expand team collaboration, add unified search, improve notifications, and continue building stronger evaluations for natural-language agent tasks.

Our goal is simple: make Ari feel like an AI teammate that is accountable for outcomes, not just answers.

Built With

Share this project:

Updates