Step 1
Upload or import your paper
Drag and drop a PDF, or paste an arXiv link or a direct PDF URL. Your document is indexed - title, authors, and metadata extracted automatically - and ready for questions in seconds.
Understand, not summarise.
A reading assistant that helps you deeply understand technical papers, right where you read them.
Used by students and researchers at Australian universities
University of Melbourne
University of Sydney
University of Western Australia
Curtin University
Monash University
The problem
You don't need another summary. You need help understanding the hard parts.
The unfamiliar term. The dense paragraph. The figure you can't quite parse.
Without draftn
PDF viewer, chatbot, notes app, search. You lose your train of thought chasing a definition across four tabs.
With draftn
Highlight the term, get an explanation at your level, in place. No extra tabs, no lost context.
Without draftn
Pasting text into a chatbot strips the document structure. Generic summaries, no citations, nothing you can verify.
With draftn
Every answer shows the paragraph and page it came from. Click through and check it yourself.
Without draftn
General AI tools are built for business reports. They oversimplify, skip equations, and ignore figures.
With draftn
Built for technical reading. LaTeX renders properly, and you can ask about any figure directly.
How it works
Read. Select. Understand.
Three simple steps to go from confused to confident.
Step 2
Select exactly what you need help with
Highlight a single term, a sentence, or a paragraph. Choose to Explain, Summarise, or Define. You control the scope - the assistant stays grounded in what you selected.
Step 3
Get answers grounded in your document
Every response is derived from the paper itself, with precise page citations you can click to jump straight to the source. Full LaTeX and markdown support for clean notation.
Features
Everything you need to deeply engage with complex material.
Selection
Select exactly what matters - in the paper or in the response
Highlight a single word, a sentence, or an entire passage in the PDF.
Choose to Explain, Summarise, or Define it - and get a response grounded in that exact selection.
The same precision extends to the assistant's responses. See a term in an explanation you don't follow? Select it and ask again. Read, ask, understand, dig deeper - without ever leaving your document.
Attention Is All You Need (2017) — Abstract
The dominant sequence transduction models are based on complex recurrent or convolutional neural networks that include an encoder and a decoder. The best performing models also connect the encoder and decoder through an attention mechanism. We propose a new simple network architecture, the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely. Experiments on two machine translation tasks show these models to be superior in quality while being more parallelizable and requiring significantly less time to train.
What are the key findings and supporting evidence?
Assistant
•
AlphaFold utilizes the Evoformer architecture to simultaneously process multiple sequence alignments (MSAs) and pairwise residue relationships, followed by a Structure Module that uses Invariant Point Attention (IPA) to predict 3D coordinates.
•
The system is trained through self-distillation — incorporating high-confidence predicted structures from unlabeled sequences back into the training set — and a BERT-style objective that masks residues to help the network learn evolutionary patterns.
•
It achieves high accuracy (median 0.96 Å r.m.s.d.95, a measure of average distance between atoms) by using the Frame-aligned point error (FAPE) loss, which ensures atoms are correctly positioned relative to local residue frames.
Citations
Every answer traces back to the source
Responses are grounded in your document, not generated from training data. Every answer includes clickable page citations - click one and the viewer jumps to the exact passage with the relevant text highlighted.
Your conversations are saved per paper. Come back days later and pick up exactly where you left off.
Technical rendering
Equations and notation, rendered properly
Responses display LaTeX, mathematical notation, and formatted tables exactly as they should appear. No broken symbols, no garbled formatting - just clean, readable answers that match how your field actually writes.
Adam: A Method for Stochastic Optimization — Section 2
As explained in section 2, Adam utilizes initialization bias correction terms. Let be the gradient of the stochastic objective , and we wish to estimate its second raw moment using an exponential moving average of the squared gradient, with decay rate . The update at timestep of the exponential moving average = · + (1 − ) · can be written as a function of the gradients at all previous timesteps:
Response depth
Cryogenic electron microscopy maps are three-dimensional reconstructions of macromolecular structures, such as proteins, created by capturing images of samples that have been flash-frozen at extremely low temperatures to preserve their native state. These maps provide a density visualization that researchers use to determine the precise spatial arrangement of atoms within a biological molecule. In this research paper, the authors note that AlphaFold's high-accuracy predictions have already assisted the experimental community in interpreting these complex maps.
Depth control
Answers that match how much you already know
Switch between undergraduate and graduate-level explanations with one click. New to the topic? Get clear foundations. Deep in your research? Get the full technical detail.
Visual understanding
Click any figure. Ask what it means.
Select any diagram, chart, or figure in the paper and ask questions about it. The AI explains what's shown - architecture diagrams, flow charts, data plots - in the context of the paper it comes from.
No more staring at Figure 3 wondering what the arrows mean.
Figure — Selected
Assistant
preview truncated intentionally for the homepage snapshot.
This diagram from the AlphaFold research paper illustrates how the accuracy of protein structure predictions depends on the amount of available evolutionary information. The y-axis measures accuracy using lDDT-Cα and the x-axis represents MSA depth (Neff).
The plot reveals a clear threshold effect...
And all the details you'd expect.
Page Focus
Focus responses on a page range for targeted answers.
URL Import
arXiv or direct PDF links
Collections
Organise by topic
Export Chat
As text file
Chat Search
Find past answers
Read Anywhere
Mobile, tablet, desktop
Auto Metadata
Titles and authors extracted
Saved Chats
Pick up where you left off
Drag and Drop
Import instantly
Use cases
Built for how you actually work.
For Students
Make sense of your readings, week by week
You've been assigned a dense 30-page paper the night before your tutorial. You don't need a summary - you need to understand the core argument well enough to discuss it.
Upload the paper, select the passages that lose you, and get explanations at a level that actually helps. Organise readings by unit. Search past conversations when exams roll around.
For Researchers
Go deeper, faster - without losing rigour
Working through a stack of papers for your literature review. Some are in your field; others are adjacent disciplines where the terminology is unfamiliar.
Import from arXiv or a direct PDF link. Use page focus to zero in on methodology sections. Toggle to Technical mode when you need the full detail. Export threads as research notes.
For Academics
Stay current across disciplines, not just your own
Reviewing a paper from an adjacent field for an interdisciplinary project, or preparing to teach material outside your core specialisation?
Orient yourself on unfamiliar terminology and methods. Toggle depth depending on whether you need a high-level orientation or a detailed breakdown. Every answer is verified with citations.
Why this is different
Not another generic PDF chatbot.
Most AI reading tools are built for anyone reading anything. We built this specifically for people who read technical, complex material and need to genuinely understand it.
Asked of “Highly fragmented European wetlands with uneven restoration needs”, Nature, 23 July 2026
Asked“What proportion of wetland area is affected by human activities, with what uncertainty, and how much soil carbon could have been lost relative to an undisturbed baseline?”
A general AI chatbot
full answer · 205 words
According to the paper, 20.4% ± 3.4% of Europe's seminatural open wetland area is affected by human activities. For carbon, the authors estimate a potential soil-carbon stock loss of 0.84–4.87 Gt CO₂-equivalent relative to an undisturbed baseline. An important qualification is that this is a stock difference, not an annual or cumulative emissions estimate.
Correct — but to check any of it you go and find the section and pages yourself.
draftn
full answer · 60 words · 3× shorter
Human disturbance: An estimated 20.4% ± 3.4% of Europe's wetland area is affected by anthropogenic disturbance; 13.6% ± 3.0% is classified as most disturbed.1
Potential soil-carbon loss: Relative to an undisturbed baseline, disturbed wetlands could have lost 0.84–4.87 Gt CO₂-equivalent — approximately 10–11% of Europe's mean wetland soil-carbon pool. This is a 2018 stock difference, not an annual emissions rate.47
Every chip jumps to that passage, highlighted in the PDF.
Pricing
Start free. Upgrade when you're ready.
Upgrade anytime. Cancel anytime. No lock-in.
Free
$0
forever
Includes:
1 paper
50 pages per document
50 messages / month
Pro
$12
USD / month
Includes:
Unlimited papers
Unlimited chats
50 pages per document
Priority support
Fair use limits apply to help prevent automated abuse and protect the experience for everyone.
Stop summarising. Start understanding.
Try it for yourself. Free to start, no credit card required.