How India’s Scroll is building a trusted workspace for the age of personal AI
This comes at a time when AI agents can already pull and synthesise journalism from anywhere.
Sannuta Raghu, who spoke at WAN-IFRA’s World News Media Congress in Marseille, laid out both the problem and Scroll’s response.
About five years ago, the organisation stopped chasing breaking news and moved toward consolidation, context, and ground reportage.
In the process, it identified its core reader: an academic or researcher who studies South Asia deeply, and who is paying for the work.
“What we’re pivoting to is turning our platform into a trusted workspace that provides an accountable environment, that creates a system of record, and provides auditable comprehensiveness,” Raghu said.
Eighteen months ago, the team began exploring whether a text article could be converted into different interfaces and formats – calculators for budget stories, knowledge graphs, mind maps.
They achieved this with high fidelity to source and high accuracy, and the result is the platform Scroll has built today, divided into three layers:
Mini, Core, and Deep: Three layers, one workspace
Mini is a structured snapshot of a story cluster – not at the article level, but at a larger event and theme cluster level. It gives the reader complete situational awareness through bullet points, key numbers, named entities, and unlimited FAQs, all personalised to a profile the reader builds and can update at any time.
Core is the article as journalism: reported, written, and owned by Scroll. “This is one space in the workspace where it’s not touched by AI,” Raghu said.
Deep is a curated dossier drawing on Scroll’s archive and algorithmically sourced trusted external material – what Raghu calls “societal information” – which can include news organisations, open datasets, and government bodies.
“We’re being very open about the comprehensiveness gap,” she said. “We’re saying Scroll is able to cover this much, and this much we are going to curate for you from trusted sources.”
The Deep layer includes timelines, knowledge graphs, gap analysis, and coverage maps, and allows readers to annotate each element within their own workspace.
Why now: The aggregation problem
The urgency behind Scroll’s pivot is straightforward. AI systems can already aggregate and synthesise news on demand.
The Nothing Phone released an AI-native operating system last year that lets users create personalised apps pulling from live news – a safety monitor for an elderly parent, a tracker for someone managing a chronic condition.
Personal agents like Open Jarvis come with news presets that run on demand, on schedule, or continuously. A newsletter-for-one, tailored to the people in a specific meeting, costs about a cent to produce.
“You can ask Claude, you can ask Open Jarvis, you can ask a whole host of these agentic systems to pull from various organisations and give you one comprehensive answer,” she said.
“So how do we deal with this as a small news organisation based out of India that is catering to academics and researchers who want an environment that they can trust?”
Events, atoms, and the system of record
Powering the platform is what Raghu describes as a system of record, built on two distinct units: events and atoms.
“An event is a real-world happening that exists independently of how it was reported,” she said.
The atom is the sentence-level layer – how a journalist explained, witnessed, and reported that event: the exact words used, who was quoted, what was attributed, and whether a sentence reports a fact or interprets meaning.
Scroll is categorising its archive, batch by batch, into events and atoms, which then powers the dynamic, personalised experiences across the platform.
The structured data underpinning this is documented at newsatom.xyz, where Raghu said there is about a hundred-page paper on the schema.
Running the entire Scroll archive of 500,000 articles through a frontier model would cost around $200,000 – impossible for a small organisation.
So Raghu’s team has built a lightweight open-source repository based on the News Atom paper that allows newsrooms to extract and test events and atoms from their own data locally, without hitting an API, at zero cost.
“I’ve tested this with IBM models, with Gemma models,” she said. “They do a pretty good job.”
