NDTV AI Strategy: Why Publishers Must Build Beyond the Models

As media organizations grapple with generative technology, NDTV Chief Product Officer Rohan Tyagi stated during the Digital Media India conference in New Delhi that publishers must shift focus from off-the-shelf AI models to proprietary workflows, software, and data layers that create enduring value.

The core challenge facing modern newsrooms is not building artificial intelligence products, but developing applications that continue to deliver utility over time. According to Tyagi, lower barriers to entry mean that deploying basic applications is no longer a differentiator in the digital publishing sector.

Instead of treating foundational models as the primary asset, media companies must own the surrounding infrastructure. This includes specialized software, operational workflows, and verified content repositories that competitors cannot easily replicate.

Evaluating Thin Versus Thick AI Layers in Modern Newsrooms

To navigate technological integration, NDTV implements an internal evaluation framework separating developments into distinct categories. Thin layers rely heavily on existing large language models acting as basic wrappers. These tools help teams move quickly, manage organizational credits, and test initial concepts.

However, over-reliance on simple wrappers risks producing what Tyagi termed vibe-coded slop—functional applications that fail to address concrete operational needs. Conversely, thick layers require deeper investments of time, money, and editorial judgment. By combining artificial intelligence with proprietary data and internal workflows, these robust systems yield long-term strategic advantages.

Constructing these advanced frameworks requires specialized engineering that standard off-the-shelf solutions cannot provide.

Automating Newsroom Workflows Without Sacrificing Editorial Oversight

Internal tool development at NDTV demonstrates how automation can streamline repetitive tasks while keeping journalists firmly in control. One notable deployment is Echion, an infographic utility that translates text articles into visual assets optimized for multilingual distribution.

Echion runs on standard off-the-shelf image models supplemented by a proprietary user interface, pre-set templates, and strict corporate branding guidelines. By embedding persistent instructions into templates, the newsroom prevents recurring software errors, such as generating incorrect geographic maps of India.

Another internal utility, Liza, functions as an agentic analytics assistant for social, product, revenue, and editorial teams. Rather than forcing staff to navigate separate tracking screens, Liza integrates directly with Google Analytics, Search Console, Chartbeat, and YouTube. Teams can query metrics and schedule regular updates directly through Discord.

Managing large-scale data ingestion and ensuring compliance across complex digital archives requires robust technical infrastructure.

Refining Consumer Experiences and Vector Search Economics

Beyond internal workflows, the organization tests AI-driven consumer features inside its primary mobile applications. Experiments include automated news feeds combining text summaries, short video loops, and animated graphics. Because early summaries produced inconsistent results, the publisher established an editorial grading interface within its content management system.

Journalists rate and review summaries to train the system, moving beyond simple binary approvals toward continuous model grading. Furthermore, the company replaced traditional machine learning recommendation models with large language model embeddings to better interpret article context.

Another consumer-facing initiative, askNDTV, acts as an answer engine drawing strictly from the company’s historical journalism and structured data repositories. To make this possible economically, the media brand had to overcome significant cost and performance hurdles associated with vector searches across decades of published text.

This technical challenge led to internal research. The team developed a novel method for structuring databases for efficient retrieval, culminating in a research paper titled Learned Rotation-Aware Binary Projections for Efficient News Retrieval, which was accepted at an information retrieval conference.

As digital archives expand, safeguarding intellectual property and proprietary codebases becomes paramount for media enterprises.

Technology alone does not guarantee editorial success, and distinct mediums require separate standards. Audio generation experiments revealed that text summaries fail when converted directly into spoken scripts without dedicated editorial intervention. By integrating cloned anchor voices with human-graded scripts, the organization built functional audio feeds rooted in verified reporting.

The evolving landscape demonstrates that sustainable advantage belongs to publishers who build resilient, proprietary layers around standard technology rather than relying on the underlying models alone. Future success depends on disciplined engineering, rigorous editorial oversight, and an unwavering commitment to proprietary data assets.