Skip to content
Atharva Palsuledesai

An AI agent that changes your plan when you get stuck

What
A personal project: a goal and task planner run by an AI agent
Built
During the CSRBOX Applied AI Internship (first version, January 2026), then audited and upgraded in September 2026
Client
None. This is my own project

The idea

A normal to-do list doesn’t adapt. If a task is too hard today, the list just sits there. OrbitAI turns a goal into a plan of 3 to 7 tasks, and when you tell it “I can’t do this today”, it changes the plan.

What it does

  • Turns a goal into an AI-written task list with priorities.
  • On each run, the agent chooses one action: suggest the next task, add tasks, reschedule a hard task (putting a small 5-minute task in its place), set a reminder, ask you a question, or just reply.
  • Refuses requests that are off-topic, such as asking it to write code.
  • Shows a progress bar, completes the goal when the last task is done, and keeps a history of finished and replaced goals.
  • Keeps a short record of the agent’s recent actions, and shows upcoming reminders.
  • Sign-up with email confirmation; light and dark themes.
  • OrbitAI dashboard on a phone
OrbitAI dashboard: the goal "Learn basic Python in 2 weeks", a progress bar at 20% and the task list
OrbitAI showing the agent's reply to a message
OrbitAI goal history with finished and replaced goals

How I kept the AI under control

  • The agent can only pick from a fixed list of actions, and must answer in a strict format. It can’t do anything the software doesn’t know how to carry out.
  • A smaller, faster model makes the frequent decisions, and a larger one writes task plans, where quality matters most. If one model fails, it falls back to the other, then to Gemini.
  • Ticking off a task doesn’t use the AI at all, so it’s instant and always works.

If you’re thinking about AI for your business, this is the approach I’d bring: the AI works inside limits the software sets, not the other way round.

What went wrong, and what I did about it

I showed the first version to friends and college professors. They suggested a progress bar, fixing errors on repeated prompts, and a way to start a new goal after finishing one. My first attempt at those changes broke the app, and with the deadline close I rolled back to the working version.

In September 2026 I came back to it with a full security audit and upgrade:

  • The cause of the earlier break. The code assumed a user has exactly one open goal, but nothing could end a goal, so starting a new one created a second open goal and the app failed. Goals now record when and how they ended, and the database refuses a second open goal.
  • A security check I ran found the database open to anyone. I switched on access rules for every table, made every request prove who is logged in, and protected the scheduled job with a secret. Then I tested it: a stranger’s request gets nothing, and fake requests are rejected.
  • The live site wasn’t updating. The hosting was holding new versions back instead of publishing them, probably since a rollback in January. I found this during testing and fixed it.

Where it stands

On 27 September 2026, all six checks passed on the live app: a stranger’s read returned nothing, unauthorised requests were rejected, the progress bar and task ticking worked, the agent replied, a new goal saved the old one as “replaced”, and finishing the last task completed the goal.

It has no user numbers, and I don’t claim any.

What’s next

  • Reminder emails don’t send yet. The free hosting plan blocks the usual email method, so this needs a different email service.
  • The link in the email still points to a local test address.
  • No password reset page, and no editing or deleting tasks by hand yet.
  • No automated tests yet.

How the work was split

  • First version (January 2026): I described what I wanted, used ChatGPT to generate the code, then tested and assembled the app.
  • September 2026 upgrade: I set the requirements (switching to Groq from Gemini, which model does which job, goal history, the progress bar) and made the decisions. I used Claude to audit the code, design the fixes and write the changes, and applied them with Claude Code. I configured the hosting and database myself, and ran all the tests.

Built with: Next.js, TypeScript, Tailwind CSS, Node.js, Express, Supabase (database and login), Groq (gpt-oss models), Google Gemini as backup, Vercel, Render.

Next: See all my work

Email meLinkedIn(opens in a new tab)