Inspiration

Most investors have a reason they bought a stock: "e-commerce keeps growing," "I'm holding this for 10 years." Then a notification lands: SHOP −9%. The app knows what you own and what happened, but it never answers the question that decides your next move: does this news actually touch the reason I bought?

We're not the only ones stuck there. In Morningstar's 2026 survey of 2,000 US investors, 54% said they felt uncertain or nervous about the market, 38% named fear of making a mistake or losing money as a top obstacle, and 24% said they struggle to know which sources to trust. Brokerages earn more when you trade, so they have little reason to build a pause between the update and the decision. We wanted to build that pause.

What it does

[NAME] links your brokerage read-only through Plaid. You write a one-line thesis for each holding: why you bought it, how long you'll hold it, and what would make you sell. After that it stays quiet. When something material happens to a stock you own, it sends one message:

SHOP −9% after earnings · your position −$140 You bought for 5+ years because "e-commerce keeps growing." You said you'd sell if "growth stops." This quarter revenue grew 24%; the drop was about margins.

Everything in italics is the user's own words from setup. The only generated content is one factual sentence about the news. It never says buy, sell or should. It puts the news next to your own rule, and you decide.

How we built it

Everything runs in n8n as a set of small workflows that share one Supabase (Postgres) database:

  1. Onboarding: Plaid sandbox /sandbox/public_token/create → /item/public_token/exchange gives an access token, then /investments/holdings/get syncs holdings and cost basis.
  2. Thesis capture: a web form posts each holding's reason, horizon and sell rule to an n8n webhook.
  3. Event detection: a schedule pulls quotes and company news from Finnhub and flags material events. A /simulate webhook lets us inject events for demos.
  4. Compose: a Code node does all the math. The dollar impact is

$$\Delta\$ = q \times p \times \frac{\%\Delta}{100}$$

where $q$ is quantity and $p$ is price, so the LLM never does arithmetic. The LLM node then writes one factual sentence as JSON.

  1. Guard: a second Code node rejects any output matching /\b(buy|sell|should|recommend)\b/i, or containing numbers that aren't in the input, and falls back to a template.
  2. Deliver: [Telegram / Twilio SMS], with daily caps and quiet hours so users don't mute it.

Challenges we ran into

  • Informing without advising. The line between "useful" and "financial advice" is thin. Our answer was to quote the user's own pre-commitments back to them instead of generating options or recommendations.
  • Keeping the AI honest with numbers. Early drafts let the model state figures, and it occasionally got them wrong. We moved all math into code and added a guard node.
  • Alert fatigue. Too many messages and people mute the bot, so rate limits and deduplication mattered as much as the alerts themselves.
  • Plaid sandbox data. Default sandbox holdings don't look like a real portfolio, so we [set up a custom sandbox user / seeded holdings] for a realistic demo.
  • [Optional, only if true: Testers found our first version's target and floor price fields confusing, so we replaced them with a plain-language sell rule.]

What we learned

  • The most valuable onboarding question is "What would make you sell?" Replaying it at the moment of a drop is more powerful than any explanation we could generate.
  • LLMs work best here as a narrow, guarded component, not the decision-maker.
  • [Add one or two of your own lessons about n8n, Plaid or teamwork.]

What's next

  • One-tap reactions (Still holding · Thinking of selling · Sold · Bought more) and a decision journal that shows when you followed your thesis and when you reacted to the price.
  • SEC filing and earnings-calendar triggers.
  • Production Plaid access, starting with Canadian brokerages.

Educational information, not financial advice.

Built With

  • n8n
Share this project:

Updates

Submission history