Scroll, an independent digital newsroom in India, saw early on that a single article format could not serve every audience need. The team had been experimenting with off-the-shelf tools that turned text stories into video summaries since 2016 but had failed to find a tool that understood journalistic values. This gap led them to build Factivo 1.0, a multilingual article-to-MP4 tool designed specifically for newsrooms to distribute short videos with high-fidelity-to-source on social platforms. 

The tool proved its value during India’s 2024 general elections, when Factivo-generated videos were piloted on Instagram. The newsroom saw a sharp spike in both followers and engagement, with audiences responding positively to the videos, which were explicitly labelled “Made with AI” and “Verified by Scroll’s Editors.”

The breakthrough: Providing information to users at the most opportune time

But Factivo 1.0 also revealed a bigger opportunity. If a text article could be turned into a video, why not also into interfaces like audio streams, calculators, decision trees, or context sliders? For Scroll, this opened the possibility of creating richer, more interactive experiences for users within their own website rather than relying on third-party social platforms.

“A text article may not always be the best way to deliver news. Some stories work better in other forms. Our challenge was: how do we give users the information they most need, in the form they most want, at the time they need it?” said Sannuta Raghu, Head of AI Lab at Scroll.

That insight became the foundation for Factivo 2.0, developed over the last nine months as part of the JournalismAI Innovation Challenge. This multi-modal versioning tool takes a single piece of journalism and transforms it into multiple formats tailored to a user’s needs.

The problem: One size does not fit all

A traditional news report on India’s Union Budget or on gold prices answers the essential questions - what happened, who said what, and when. While this context is crucial, readers often also want to know something more immediate: what does this mean for me?

For example, alongside a budget story, some readers may want only the Finance Minister’s soundbites. Others might prefer a calculator that shows their exact tax liability. With a gold price update, users may want to instantly see how much they could buy or sell gold for that day. Scroll wanted to not only explain events in context, but also give users tools to directly understand their personal impact. Repurposing every story into such tailored formats manually was impossible for a small newsroom. AI offered a way to bridge this gap at scale.

Building the solution: One story, many hats

Currently live as a public sandbox, a user is able to access a news report as summary sliders, timelines, mind maps, decision trees, calculators and expanders, at the click of a button. Each form-to-form versioning of Scroll’s articles with Factivo 2.0 is executed with large language models (LLMs). It is dynamic, based on verified news reporting and is designed to have high fidelity to source. 

In addition to these formats and their contextual variations, Scroll also developed a lo-fi news stream player and workflow. The idea was to provide an audio stream of news that users could listen to while performing routine chores - like folding clothes or washing dishes, if they wanted to. The audio is neutral-toned, non-alarmist and has an optional lo-fi background track. 

The lo-fi news stream is borrowed from Scroll’s news feed, The Latest. Every article published on The Latest, is summarised, voiced and sent to the playlist. The workflow is designed such that every story is linked to the previous one in the stream - making it seem like a long radio programme. “This format is still a work in progress but the proof of concept has given us confidence.”

Under the hood: How vectorisation helped 

One of the breakthroughs of Factivo 2.0 was vectorising Scroll’s news archive. In simple terms, vectorising means turning every article - and even every sentence - into a numerical “fingerprint” that captures its meaning. These fingerprints live in a database, where similar ideas, events, or entities can be instantly found and connected. Unlike keyword search, which only looks for exact words, vectorisation understands the context and relationships between pieces of information.

Vectorisation helped Scroll unlock their archive and extract facts from temporally and semantically similar events. “If we have produced 20 reports on a particular flood in 2025, the timeline or expander or mindmap format is able to accurately extract facts from these 20 reports, and not from reports about other floods, or a previous flood in the same geography. Our users had been asking for deeper engagement with our archives, and this is a great way to meaningfully deliver that,” Raghu said. 

Team structure and challenges faced

The team that is building Factivo is organised in three layers: Raghu and a machine learning engineer are building, experimenting, testing, debugging and iterating everyday. This forms the first layer. Scroll’s chief product officer, chief architect and editor-in-chief come in to provide feedback and guardrails. They are constantly updated on Slack everyday. This forms the second layer. The third layer is the designers and editorial testers, who come in as and when required. Scroll’s team is completely remote and works across two timezones.

Though working async has not yet proved to be a challenge, the team has faced several challenges while building Factivo.

Adapting the formats to different Indian languages continues to be a problem. Indian languages are codified as accurately as English yet. “How do you deliver the calculator dynamically with Hindi, for example? This is what we've been testing and the results are usually not precise. So we’re not able to automate it,” she added.

Another significant challenge was ensuring temporal precision in archival retrieval to ensure context remains relevant to the specific story - a problem they worked on for three months.

Versioning of quotes, especially complex or nested quotes was another challenge while working with formats that needed rewriting (like the complexity slider or the lo-fi news stream). By default, many quotes used to get summarised as statements and that needed specific prompting to be corrected.

Lessons for newsrooms

• AI as a force multiplier: Recognise that AI can be a force multiplier, unlocking potential and capabilities that would be very difficult for a small, resource-strapped team to achieve otherwise.

• Ambitious products with tiny teams: Believe that ambitious, high-quality products can be built by a tiny, committed, and skilled team. “I am not an engineer, I am a journalist - I have learnt everything on the job. From basic coding to using the same jargon as our engineer. And the same goes for him: He has understood how to teach journalism to models, “ Raghu said.

• Deeply integrate editorial input: Avoid building in silos. Work deeply with editors and journalists because the real challenge in using AI is not the technical coding, but the deliberate design required to teach a model how to preserve the original editorial intent of journalistic work.