← All projects
Live Featured

Pulse

A note-taking stream that remembers for you

Write in a single, endless stream. While you type, Pulse quietly brings back what you already wrote on the same topic, using semantic search instead of folders and tags.

Role
Builder
Year
2026
Status
Live
Read
2 min
Stack
Next.jsConvexVector SearchTipTapTransformers.js
The story

The problem

Folders, tags and graph views turn note-taking into admin work. The act of thinking gets buried under the effort of deciding where each note should go.

The approach

A single chat-like stream with no folders. While you type, the text is turned into embeddings and matched against past notes, so related thoughts surface on their own.

My role

Concept, interaction design and build, using AI to explore the trade-offs of semantic retrieval.

Where it stands

Live and tested on my own daily logs. The retrieval works; the experience still needs tuning.

Next step

Make suggestions 'polite': show fewer and better matches, give more weight to recent notes, and add an AI summary layer for reviewing the stream.

The project

The idea

Pulse came from a frustration with traditional note-taking apps. Every tool asks you to organise first and think later: which folder, which tag, which page. I wanted the opposite: a tool that never asks where a note goes, and instead tells you what you’ve already thought about the topic.

The picture in my head was a chat with my own past self. One infinite vertical stream, where every entry is a heartbeat, hence the name.

What it does

  • One stream, no structure: you just write, like in a chat.
  • Context while you type: every few keystrokes the current text is compared with all previous notes, and the closest ones appear next to the editor.
  • Search without searching: you don’t type queries. Related notes find you while the thought is still forming.

How I built it

Before writing code I used AI to explore the hard questions of semantic retrieval:

  • Local or cloud embeddings? Running models in the browser with transformers.js keeps notes private, an API is faster to set up.
  • Signal or noise? How many matches are useful before suggestions become a distraction? AI helped me simulate scenarios before building them.

The build pairs a custom TipTap editor, tuned for a chat-like flow, with Convex as backend. The text being typed is debounced, vectorised and matched with Convex Vector Search against thousands of previous entries. Matching takes well under a second, so suggestions appear while the sentence is still being written.

What I learned

Testing it on real daily logs was humbling. At first Pulse showed matches for every single sentence, and it felt like someone whispering in my ear while I was trying to talk. The key lesson: semantic relevance is not cognitive relevance. Two notes can be similar without being useful to each other in that moment.

I added “soft context” rules: higher similarity thresholds and more weight to recent notes. I also realised that an infinite stream is great for capturing thoughts, but poor for reviewing them without a summary layer on top.

The idea holds. The interface has to be polite: present when you need it, invisible when you don’t. It’s a tool for non-linear thinkers, and it still needs a few linear constraints to be useful.

Other experiments
Got an idea?

Let's talk.

Collaborations, product challenges, or just a few thoughts on AI and delivery.

Contact me →