Notes You Did Not Write A Browser That Annotates the Page Before You Ask

This application covers a web browser that reads the page you have just opened, compares it mathematically against a library of other pages, documents and human written annotations, and then places its own annotations on the page for you. The applicant is Sunil Pinnamaneni, an individual inventor whose earlier work in this area was commercialised as the ExactNote browser extension. The claims run from a computer implemented method through computer readable media to a full system.

The Problem

The background section starts from ordinary web behaviour rather than a technical gap. People are spending more time on the web than they used to, and as they do, they increasingly need to retrieve something they saw earlier, share a snippet with a colleague, ask a question about part of a document, connect information across several documents, judge how trustworthy a crowd sourced claim is, and read other people’s comments in the context where they were made.

The specification then lists what people actually do instead. They repeat searches to find content they already read. They rescan articles for a half remembered sentence. They copy and paste sections into emails. They refer to text by paragraph number or page position while talking on the phone. They paste a quote into a comments section so their question has context. The document calls this inefficiency and time waste for a large number of users, and adds a second point: the absence of tools for flagging inaccuracies has allowed wrong information to persist longer than it should.

What This Invention Does

The heart of the application is web annotation as a first class browser feature, with machine learning bolted on so that annotations can be generated automatically.

The manual half comes first. A user highlights text and attaches a comment, and the pairing is tagged with what the specification calls an annotation semantic relationship: question, disagreement, agreement, exclamatory or advertisement. Each relationship has its own small symbol, and the symbols are what other users see inline, so a reader can tell at a glance that someone questioned a sentence or disputed it without opening the comment. A design constraint stated repeatedly is that the document viewing context should be modified as little as possible. Annotations can be gathered into collections and shared, and the worked example is a teacher who builds an American Presidents study guide from question annotations left on Wikipedia pages and shares it with a class.

Claim 1 describes the automatic half as a pipeline. A source set of sentence embedding vectors is calculated for sentences drawn from a body of source content, which may be web pages, documents, or annotations of them. When the user navigates to a URL and the browser loads a document, a request set of embedding vectors is calculated the same way, this time from the document itself, its annotations, replies to those annotations, and any items in a collection containing the document. The two sets of vectors are used together to retrieve, through APIs, a subset of the source content. A deep learning model for natural language processing then auto-annotates part of the request content using what was retrieved, and the results are displayed on the page in the browser.

The dependent claims fill in the machinery. Auto-annotations can be items containing links to similar content on other documents, augmented with generated text summaries of the linked articles, and claim 7 specifies a deep learning model with a transformer architecture. Claims 8 to 14 restate the method as one or more non-transitory computer readable media, and claims 15 to 21 restate it as a system.

The description is more specific than the claims. A bidirectional long short-term memory model with a final max pooling layer is offered for computing the sentence embeddings. A convolutional neural network checks whether the relationship a user chose actually matches what they wrote. BERT handles named entity recognition inside pages, so that entities such as Wikipedia subjects or products for sale can themselves be auto-annotated, and a BERT model fine tuned for sentiment analysis flags inappropriate user generated content. Advertising gets its own annotation relationship, positioned for contextual relevance and revealed in full only once the reader acts out of curiosity.

Key Features

• Annotations that carry a relationship, not just a comment. Every highlight and comment pair is typed as a question, disagreement, agreement, exclamation or advertisement, and each type has its own inline symbol.
• Meaning compared as vectors. Sentences from the page, its annotations, replies and collection are embedded and matched against an embedded library of source content, so similarity is semantic rather than keyword based.
• Retrieval followed by generation. Related material is pulled through APIs first, then fed to a language model that writes the annotation, a structure that prefigures what is now called retrieval augmented generation.
• Auto-annotations that link and summarise. The generated annotations carry links to similar passages elsewhere, augmented with machine written summaries of those articles.
• Minimal disruption to reading. An unchanged document viewing context is treated as an objective in its own right, with symbols rather than text boxes carrying the signal.
• Advertising as an annotation type. Advertisements run through the same mechanism, positioned for contextual relevance and revealed in full only after the reader acts.

Who Is Behind It

The applicant is an individual rather than a company, which is unusual for software of this scale. Sunil Pinnamaneni and co-inventor Rona Sfakianakis are the pair behind ExactNote, Inc., which shipped a Chrome and Firefox extension for creating, viewing and organising annotations on web pages and PDF documents. The company’s own website no longer resolves, so the surviving public record is the extension listings and the patent family.

That family starts with United States application 16/679,278, filed on 10 November 2019 and granted as US 11,321,515. International application PCT/US2020/059750 followed on 9 November 2020, published as WO 2021/092592. Its Australian branch is application 2020378213, and the present application is a divisional of it, opening with the standard formula incorporating that parent specification by reference.

Why It Matters

The dates are the most striking thing about this document. The priority work was done in 2019 and the international filing landed in November 2020, well before large language models became a consumer product. Yet the pipeline in claim 1, embed a corpus, embed the current context, retrieve the nearest matches through an API, hand them to a language model, display generated text back to the user, is recognisably the architecture the industry has since converged on. Filings from that window are worth attention precisely because they describe now standard patterns in the vocabulary that existed before those patterns had names.

The second point is about the annotation layer itself. A shared, typed, machine readable annotation layer over the web has been attempted repeatedly, and it has never quite reached escape velocity because the value only appears once enough people annotate. This application proposes a way around that bootstrapping problem: if the system can generate credible annotations from existing content, the layer is populated before any user contributes to it.

Finally, the filing strategy. Holding the application in the inventor’s own name rather than the company’s keeps the asset independent of the corporate entity, which matters when the operating company has gone quiet, and a 2026 divisional off a 2020 parent keeps a claim drafting option open in a market where examination of software subject matter has been notably strict.

Related Concepts

• Web annotation – the layer of user commentary over web documents that the whole system is built around.
• Sentence embedding – the vector representation used to judge whether two passages mean the same thing.
• Transformer architecture – the deep learning model type named in the dependent claims.
• Retrieval augmented generation – the retrieve then generate pattern that claim 1 describes in earlier language.
• Named entity recognition – the BERT based step that finds the people, places and products worth annotating.
• Hypothes.is – the best known open attempt at a shared annotation layer for the web.

---

AU 2026202105 was published in the Australian Official Journal of Patents on 16 April 2026 and is open for public inspection. Patent applications represent inventions that are sought to be protected and do not necessarily reflect commercially available products.