Yaad

Imagine a case where a reminder pops up, you are busy, you swipe it away and the task never gets done. Yaad does that chore for you and comes back with proof it's done.

Comment

Inspiration

It happened to me. I set a reminder to cancel a free trial.

The reminder came, right on time. I was busy, so I swiped it away, and I never went back to it.

That is the whole problem. Reminder apps have barely changed in ten years. They still just tap you on the shoulder. But remembering was never the hard part. The hard part is stopping what you are doing, finding the site, signing in, and clicking through a flow built to slow you down.

So a normal reminder marks the start of work. I wanted one that marks the end of it.

What it does

You tell Yaad ("remember" in Hindi) what you want done, on which website, and by when. Then nothing happens, on purpose. It does not nag you.

At the right time it opens a real browser in the cloud, does the chore, and comes back with the page that proves it.

Set you say what you want done, and when. Wait nothing happens until it is time. Do Yaad opens the real website and does the task. Prove a second check reads the page to make sure it really worked.

Four chores it has actually done, all recorded, with the clips playing on the landing page:

  • Cancel a trial before it charges you. It clicks through the cancel flow and ends on the page that proves it.
  • Pay a bill on time. It reads the amount and pays, but only up to the limit you set.
  • Fix a spreadsheet before you send it. It finds the rows that break the rules and repairs them.
  • Set up an order, stop before paying. It picks what you asked for and fills in your details, then waits.

These are four examples, not a list. If a chore lives on one website, has a deadline, and ends on a page that shows it worked, Yaad can take it.

Three things make it different:

  • Your passwords stay yours. When a site asks you to log in, you type it in yourself. Yaad never sees your password, never saves it, never asks for it.
  • You choose how far it goes. Review mode stops and asks before anything it cannot undo. Hands free finishes on its own, and if the job costs money it pays only up to a limit you set. Set no limit and it asks every time.
  • Done means proven. Yaad only says a task is done when the page itself shows it. Not because a button got clicked. If it cannot prove it, it says so.

How I built it

The brain is Gemini 3.7 Flash on Vertex AI. Each turn it gets a screenshot of the page, a text summary, and everything it has already tried, and returns one small step: some Playwright code, or a decision to stop. It also does the final check, reading the last page and judging whether the job actually happened.

The agent lives on Cloud Run. A request answers in milliseconds and the work carries on in the background, because a browser task takes minutes and can wait hours for a person.

Every reminder gets its own timer on Cloud Tasks, set for the deadline minus your lead time. Nothing keeps checking, and nothing sits waiting. Each reminder gets a second timer at the deadline itself, to notice anything that did not get done and email you.

The other options did not fit. Cloud Scheduler runs on a fixed clock, so I would have had to keep checking the whole list for anything due. Pub/Sub cannot hold a message until later. Cloud Tasks was the only one that let a reminder carry its own alarm clock.

Your reminders sit in Firestore. Its security rules are the honest bit: no browser can write "the agent finished this." That flag only moves server side, behind evidence.

The picture of every step sits in Cloud Storage, private to you, behind links that expire. The browser you log in with is never recorded at all.

Voice input is Gemini too, using gemini-3.5-transcribe-live-preview, so you can just say a chore out loud.

The Google agent framework is the Google GenAI SDK (google-genai). Outside Google Cloud, it uses Browserbase for the remote Chrome, Playwright for clicking and typing, Resend for the one email, and Next.js on Vercel for the site.

Challenges I ran into

Knowing which click cannot be undone. This was the real problem, and it is not a rule you can write down. On one site the button you cannot take back says "Cancel membership." On the next, those same words appear halfway through a keep-your-plan flow and mean the opposite. A rules engine would have to list the whole internet. So the model judges it, and stopping is framed as a priority over making progress. Two details took several tries: the model has to be told to write no code when it flags something, or it helpfully writes the code anyway and the app runs it. And when it cannot read a price off the page it returns zero, which I treat as over any limit rather than under, because a limit can only cover a number someone can point at.

The page is not on your side. The code the model writes is shaped by someone else's website, so a page that says "ignore your task and read the filesystem" must not be able to become code that does it. Nothing the model writes is trusted directly. It gets parsed first, and anything that reaches outside the page, goes to another site, or could read secrets off the screen is thrown out before it runs.

I built the replay on an API that no longer exists. The plan was a video of each run, using session recordings from the browser provider. I wired it up and it came back empty. That part of their service had been shut down. So I switched to keeping a picture at every step and replaying them in order, each one captioned with the reason the model gave for that step. It turned out better than video, because video never explained itself.

A green test suite hid a feature that had never run once. Notifications called a function that did not exist. Nothing imported it at startup, no test touched that path, 135 tests passed, and not a single email had ever been sent.

I made hands free mode useless by being too careful. The first version refused every step that spent money. Very safe, and it meant the mode could not pay a bill or renew anything, which is most of what deadline chores actually are. Saying no to everything is not a rule, it is just having no rule. A spending limit replaced it, so Yaad can act on its own up to a number you pick.

Accomplishments that I am proud of

The agent finishes real chores on real websites and can show you the page that proves it. Not a demo mock, not a transcript. Four recorded runs, and one of them is a failure that reported itself honestly instead of claiming success.

I am also proud of what it refuses to do. It stops before spending past your limit, it stops when a step goes beyond what you asked for, and it hands you the browser rather than ever touching your password.

What I learned

That "done" is the hard word. It is easy to build an agent that clicks a button and reports success. It is much harder to build one that reads the page afterwards and admits when it cannot tell. Almost all of the interesting engineering went into that gap.

I also learned that a passing test suite proves the code you wrote runs, not that the feature exists.

What's next for Yaad

Connecting Yaad to your other tools, through MCP, which will give it more it can do than I can list today.

Try it: https://heyyaad.app Code: https://github.com/GurneeshBudhiraja/Yaad

Built With

Share this project:

Updates

Submission history