Inspiration ## Inspiration

What I know is scattered: lecture PDFs, YouTube videos, links, plans, half-written notes. AI could help me use all of it, but most tools want everything sent to one cloud model. I wanted a notebook that understands what I put in it, where a small model decides what each request needs and private notes never leave for the cloud.

## What it does

Second Brain stands on three pillars:

| Pillar | What it means | What the app does | |---|---|---| | Store | Keep what you learn, in a structure that lasts | Blocks and databases; sources (PDF, web, video, audio) kept beside the notes made from them; a privacy level on every note | | Link | Connect each new thing to what you already know | Links and backlinks; database relations; every AI claim cites the exact page or moment it came from | | Analyse | Turn it into understanding and action | Nemotron study guides from your sources, answers grounded in your notes, interactive blocks built on request, web research |

In the demo, one study session goes from an empty note to a filed lecture:

A note is made of blocks: headings, to-dos, callouts, toggles, formulas, tables, code.
In the workspace, I add a 3Blue1Brown lecture beside the note. Nemotron reads the transcript and writes a study guide into the note.
The transcript follows the video. One click puts a frame or a clip of the lecture into the note.
Nemotron Super builds a small app that runs inside the note: a gradient-descent playground for the loss $$C(w) = (w - 2)^2, \qquad \frac{dC}{dw} = 2(w - 2), \qquad w \leftarrow w - \eta \frac{dC}{dw}$$
When the lecture isn't enough, it searches the web with Tavily, and I approve the exact words that leave.
The note lives in my course database, as a row with properties that opens as a full note.

## How we built it

Frontend: Next.js and BlockNote on Vercel. Backend: FastAPI on Render. Data: Supabase (Postgres and pgvector).
Nebius Token Factory serves every AI call of the demo account:
- NVIDIA-Nemotron-3-Nano-30B-A3B is the brain: it reads each request and decides where it runs, and it answers everyday questions.
- nemotron-3-super-120b-a12b does the heavy work: study guides, derivations, interactive blocks.
A switchboard runs before every model call: fixed rules first (private notes go to the local model only), then the brain, then a slot. Every answer shows which model ran and why, and Settings → Routing logs every call.
Tavily handles web search, behind an approve-first step.
The demo video is made from screen recordings of the real app, edited with Remotion.

## Challenges we ran into

YouTube blocks cloud servers. Captions and media requests from Render are refused ("Sign in to confirm you're not a bot"). Now a subtitle file in the session gives a YouTube source its transcript, and frame and clip capture work when the app runs on a home connection, which is how the demo was recorded.
Nemotron has its own habits. It writes links in 【】 brackets, sometimes double-escapes newlines, and streams formulas half-finished. Each of these got a small fix and a test.
A 4 GB laptop. Rendering the video stalled until we converted the recordings to a constant frame rate and rendered in short chunks.

## What we learned

A small model makes a good switchboard: it decides in about a second, so the big model is only paid for when the work needs it. And privacy is better as a rule in code than as a line in a prompt.

## What's next

The plan grows the three pillars in five phases:

Foundation (built): blocks and databases, the knowledge substrate.
Agency (built here): routing and privacy rules, skills, MCP tools, and a learning loop that drafts skills you approve.
Reach (started): capture all your digital life as you consume it, including social media, newsletters, messages and podcasts.
Ownership (idea): a small portable device that runs the small Nemotron and keeps your private memory.
Presence (idea): smart glasses, wearables and sensors, processed on your own device under the same rules.

Before the brain sees your real life, it has to run on hardware you own.