As media organizations rush to capitalize on generative artificial intelligence, Rohan Tyagi, Chief Product Officer at NDTV, has a blunt warning: most publisher AI products end up abandoned once the initial novelty fades. Speaking at the Digital Media India conference in New Delhi, Tyagi argued that the plunging entry barrier for building basic software means the AI models themselves are no longer a competitive advantage. Instead, the real challenge for modern newsrooms lies in what publishers build around those models to secure long-term value.
Rather than treating generic large language models as the final product, NDTV’s product strategy separates AI integration into two distinct tiers: thin wrappers and thick, proprietary layers. According to Tyagi, publishers must shift their focus toward owning the underlying software, customized workflows, and proprietary data structures that competitors cannot easily replicate.
Evaluating AI tools through this strategic lens allows newsrooms to avoid what Tyagi termed “vibe-coded slop”—hastily assembled AI features that fail to solve actual user problems. By distinguishing between quick-fix interfaces and deeply integrated technical architecture, the broadcaster aims to build scalable solutions that withstand shifting industry trends and deliver lasting utility to both journalists and audiences.
Building AI for Newsroom Workflows
Internal tool development at NDTV focuses on solving concrete newsroom friction points while keeping editorial staff firmly in the loop. One prominent example is Echion, an automated infographic tool designed to convert written articles into multi-language visual graphics optimized for digital publishing.
“It’s mostly off-the-shelf image model underneath, with a UI on top,” Tyagi said during the New Delhi conference, noting that the publisher invested heavily in pre-set templates, standardized branding guidelines, and default prompts.
By embedding strict visual parameters into templates, the newsroom eliminates recurring generative errors, such as AI models repeatedly producing incorrect maps of India. Instead of forcing journalists to manually correct repeated prompts, the system applies predefined constraints automatically. Another internal creation is Liza, an agentic analytics tool that pulls data from Google Analytics, Search Console, Chartbeat, and YouTube. Rather than forcing teams to navigate separate dashboards, Liza answers natural-language queries and dispatches recurring reports directly through Discord.
“This is actually one of the only real value agentic AI products I have seen internally,” Tyagi said.
Reimagining Consumer News Experiences
Beyond internal production, NDTV is testing AI across its consumer-facing applications to alter how audiences discover and consume journalism. Inside the publisher’s mobile app, developers are experimenting with AI-generated news cards that synthesize articles using summaries, short videos, and animated GIFs.
Early automated summaries produced inconsistent results, prompting the creation of an internal content management system playground where editors review, grade, and correct AI outputs. Rather than relying on simple approval or rejection buttons, the publisher uses human feedback to train grading models that continuously refine future outputs.
The company is also testing recommendation systems powered by large language model embeddings rather than traditional machine learning models. These advanced embeddings better understand the contextual nuances of articles, while separate experiments automatically cluster stories around specific entities to streamline news navigation.
Another major consumer feature is askNDTV, an answer engine built directly on the publisher’s extensive digital archive. Unlike standard keyword searches that retrieve flat links, askNDTV is engineered to answer user queries using verified journalism and structured datasets, ensuring responses remain grounded in trusted reporting. Additionally, the newsroom is experimenting with AI-generated audio experiences, utilizing cloned anchor voices and algorithmic recommendations. During development, producers discovered that written summaries required significant editorial adjustments to function effectively as natural-sounding audio scripts.
Investing in Proprietary Retrieval and Thick AI Layers
Scaling consumer products across decades of published archives exposed severe performance and cost limitations inherent in standard vector searches and LLM embeddings. To overcome these technical hurdles, NDTV’s engineering teams invested in advanced retrieval research to optimize data handling across millions of articles.
The publisher developed a novel database structuring method specifically designed for efficient LLM retrieval. This research was detailed in a paper titled Learned Rotation-Aware Binary Projections for Efficient News Retrieval, which was accepted at an international information retrieval conference. Tyagi pointed to this technical breakthrough as a prime example of a “thick AI layer”—an intensive upfront investment that takes considerable time and resources but grows increasingly valuable over time.
As publishers navigate the evolving media landscape, NDTV’s approach highlights a clear operational divide. While basic AI wrappers offer rapid deployment, enduring publishing value requires proprietary data integration, rigorous editorial grading models, and heavy investment in foundational software architecture.
Worth a look