Study guides built from your own source material

A whole chapter, distilled into a guide you actually reopen.

Upload your textbook, point at a chapter, and get a print-ready pocket guide: the intuition in plain language, the technical vocabulary in the language of the field, the formulas right, and the classic traps marked.

Access is invite-only for now.

PART 1 — SHRINKAGE

1Ridge regression (L2)AI Engineer

PRACTICAL INTUITION

Ridge is a progressive tax: every coefficient pays something, the biggest ones pay most, and nobody gets fired. Everyone stays on the payroll at the end of it — just on a smaller salary.

  • Penaltyadds the squared L2 norm of the coefficients to the loss.
  • Effectshrinks coefficients toward zero, but never exactly to zero.
  • Tuningone knob only — the penalty strength, chosen by cross-validation.
RSS(β) + λ Σ βⱼ²
KEY TAKEAWAY

The same mechanism carries two names: ridge in statistics, weight decay in deep learning. The update multiplies the weight by (1 − ηλ) at every step — hence "decay". Seeing that they are one idea halves the studying.

Cross-link: tree pruning’s complexity parameter plays the same role as the Lasso penalty — a tax on complexity, in a different currency.

A real fragment from a generated guide — Regularization, bilingual mode.


The problem

Rereading a summary feels great and works badly.

On the second read the text feels familiar, and the brain quietly swaps familiarity for mastery. You close the book sure that you know it, and find out in the exam that you recognised it without being able to retrieve it.

What the learning evidence actually supports is different: pulling things out of memory (the testing effect) and spreading reviews over time (the spacing effect). A good summary is not the end of studying — it is the reference layer everything else stands on.

Three layers, three jobs

1

Reference

The pocket guide. Read it once without trying to memorise, and come back to it when you get stuck.

2

Retrieval

Quiz and flashcards drawn from the guide itself, weighted toward what you got wrong.

3

Spacing

Anki's job, not ours. Its algorithm beats anything we would hand-roll.

How it works

Four steps, and you only see one of them.

1

You point at the chapter

Upload the book — 300 pages, 800, it makes no difference — and give the range: section 7.1 through 7.6. Only that slice is read. The rest of the book is never looked at.

2

The guide is planned before it is written

First comes a plan: which thematic blocks, how many sections in each, which concept lands where. Then each block is written separately against that plan. That is what stops a twenty-page document from drifting halfway through.

3

Twelve automatic checks

Does every section open with an analogy? Is any intuition box smuggling in a formula? Are the theory boxes teaching, or just restating the answer? Every rule in the blueprint becomes a mechanical check, and the result ships next to the guide.

4

An A4 PDF comes out

Formulas properly rendered, page breaks controlled, ready to print and scribble on. Not a screen you scroll — paper.

Anatomy of a guide

Every box earns its place.

The format is not decorative. It is a written specification with a usage rule for each element, and meeting that specification is the generator’s actual job.

PART 1 — SHRINKAGE

1Ridge regression (L2)AI Engineer

PRACTICAL INTUITION

Ridge is a progressive tax: every coefficient pays something, the biggest ones pay most, and nobody gets fired. Everyone stays on the payroll at the end of it — just on a smaller salary.

  • Penaltyadds the squared L2 norm of the coefficients to the loss.
  • Effectshrinks coefficients toward zero, but never exactly to zero.
  • Tuningone knob only — the penalty strength, chosen by cross-validation.
RSS(β) + λ Σ βⱼ²
KEY TAKEAWAY

The same mechanism carries two names: ridge in statistics, weight decay in deep learning. The update multiplies the weight by (1 − ηλ) at every step — hence "decay". Seeing that they are one idea halves the studying.

Cross-link: tree pruning’s complexity parameter plays the same role as the Lasso penalty — a tax on complexity, in a different currency.

Thematic block

Groups sections by subject. Two to four per document.

Intuition

Opens every section, before anything technical. Always a concrete analogy, never a formula, two to four sentences.

Technical content

Definitions, notation and formulas in the language of the literature. This is what you will meet again in the paper.

Deep dive

Only for the concepts that genuinely stick. Overused, the box stops meaning anything, so it has a ceiling: at most one every two sections.

Cross-module link

Ridge in statistics is weight decay in deep learning. Seeing that they are one mechanism halves the studying.

Language

Three modes. None of them is a translation.

An analogy is born in a language, not translated into one. Asking for the same chapter in another language regenerates it from the source: more tokens, and the only version that works.

Bilingual

default

Headings, technical terms, definitions and formulas in English. Intuition, deep dives and traps in your own language. This is the reference format.

100% English

Everything in English, analogies included. Not one word of the other language survives.

100% Portuguese

Everything in Portuguese. The canonical English term appears once in parentheses at first mention, then never again.

The method

The tool is one piece. The loop is what works.

  1. 01Generate the guideRead it once, end to end, without trying to memorise anything.
  2. 02Close it and answerA retrieval quiz, from memory. Note where you stalled.
  3. 03Turn gaps into cardsFlashcards pulled from the guide, weighted toward what you missed.
  4. 04Let Anki do the spacingSpaced repetition is its job. We export and get out of the way.
  5. 05Build somethingOnce the concepts pile up, stop summarising and make something with them.

Under construction

Pocket guides, exam reviews, the retrieval quiz and flashcards with Anki export all work today — inside every generated guide. Tracking concepts across guides to time the anchor project is what comes next.

Plans

Start free. Pay only when you want the complete study experience.

Every month you can create one PDF guide for free. Plus includes four complete guides with quizzes, flashcards, projects and Anki export.

Free
USD 0

1 PDF guide each month

Forever. Created with Claude Sonnet 5, without quizzes or flashcards.

Plusbest experience
USD 9

4 complete guides

Claude Opus 5, with quizzes, flashcards, projects and Anki export. Buy extra credits whenever you need them.

Plus credits never expire. The platform remains invite-only; checkout is coming next.

Questions

What people ask.

Do I have to upload the whole book?

You can. But give the section range: only that slice is read, and it cuts the cost of a document by roughly three times.

Does it work with scanned PDFs?

No. If the PDF is images there is no text to extract, and the platform refuses rather than inventing content. Run OCR first.

Is the content made up?

Everything comes from the slice you uploaded. Where the source is ambiguous or a section is missing, the guide says so instead of filling the hole.

How long does it take?

A few minutes per chapter. The screen shows the stage: planning, writing block 2 of 3, rendering.

Can I get the same chapter in another language?

Yes, and it is generated from scratch rather than translated. It costs a new credit, for the same reason.

A whole chapter, distilled into a guide you actually reopen.

Sign in

Access is invite-only for now.

summario — study guides from your own material