Innovating at Speed in an Underrated Drug-Making Process

By Deborah Borfitz 

October 22, 2026 | Novartis recently leapt to the front of the pack with a self-service platform for digitalizing chemical manufacturing controls (CMC) built on a new-to-market enterprise solution from Genedata that leverages artificial intelligence (AI) agents providing a “Gemini-like” experience to users. The collaborative feat seamlessly connects data and workflows across modalities and sites, significantly boosting operational efficiency and submission timelines for Novartis. It also earned the company an Innovative Practices Award at the Bio-IT World Conference and Expo earlier this year in Boston. 

Genedata Vico launched commercially at the event, and the journey of Novartis as an early adopter and co-developer of the platform (known internally as Scientific Product Information Network Exchange, or SPINE, Explorer) for enabling in silico experimentation and use cases such as real-time monitoring of the drug manufacturing process was retold in a presentation by Ganga Kalidindi, global head of data and AI for technical research and development. This is a use case of preventing batch losses, which is but one of endless possibilities to automate and optimize CMC processes, he says.  

Although the value of CMC is often underappreciated by the pharma industry relative to research, says Kalidindi, it “ensures quality, safety, and the efficacy of a drug, [and] directly impacts regulatory approvals.” Roughly three-quarters of all complete response letters issued by the U.S. Food and Drug Administration (FDA) cite CMC, quality, or manufacturing deficiencies. 

Given the abundance of data available within Novartis, it seemed logical that it could be harnessed in some way to accelerate its time-to-IND (investigational new drug) timelines. But the data was in the form of scanned PDFs, and well over 4 million of those paper documents were being produced annually. “It was just a huge dump of data, all over the place,” he says, with each acquisition by the company scattering more data with its own jargon and in its own document format across siloed systems. “There was no semantic orchestration or alignment at all.”  

Before November 2022, when ChatGPT accelerated the AI boom, data scientists were just putting data from the scanned documents into spreadsheets internally to glean some insights, says Kalidindi. The mission was to better organize the data to optimize decision-making by transforming Novartis from a document-centric to a data-centric organization.   

Starting Point 

Getting all the data organized in one place was to allow smart people (scientists) to ask the right questions, says Kalidindi, but AI was not part of the picture at the starting point with SPINE Explorer. The focus was instead on semantically organizing the data around an ontology, building knowledge graphs, and specialized databases known as triple stores.  

The journey began with an off-the-shelf enterprise software-as-a-service platform known as Genedata Profiler, since it provided all the basics: centralized repository, access controls, data and algorithmic fallibility (now with the help of an AI annotation assistant), and it operates as a GxP-compliant system that ensures strict data security (no data leakage via AI, for example) and quality. Moreover, any tool can access data in the lakehouse. Novartis opted not to “double up” with a custom solution that often translates into technical debt, Kalidindi says.   

The concepts of “data mesh” and “data products” were then becoming popular, he adds, the core idea being to move away from old, centralized data teams toward decentralized ownership. In parallel, generative AI was emerging with capabilities that included writing code.  

Project kickoff was marked by digitalization projects to get rid of as much paper as possible, says Kalidindi, followed by inputting the data into Genedata Profiler with naming standards and comprehensive semantic orchestration. During this intake process, the semantic correlation between data coming from different sources became clearer, and subject matter experts were involved with metadata management. 

“We captured all the business rules, the metadata, the reference list, [and] the ontology, into a separate system and that is where we did data governance,” he says. “We were not trying to structure data at the source ... that would have been a boiling-the-ocean kind of approach ... This data is for secondary use.”  

Kalidindi’s team at Novartis integrated AI models into Genedata Profiler, initially to assist data scientists in writing code and later to help meta-tag the data, continuously refining the data as it came into the platform. Within the platform itself an AI agent could annotate data on the fly, he adds, using approved annotations as a starting point for new suggested annotations. 

Given a model, Kalidindi explains, his team can build some ontologies on the side that the AI agent can use to inspect and organize data and build data products packaged to solve a specific business problem or user need. If there is no model, the agent can build one, which, once approved by a human-in-the-loop, can be leveraged to build a data product. 

“The platform is capable of solving the hardest problem, which is how ... [to] clean, curate, harmonize, and build a data product out of data, all leveraging AI,” he says. It also turns data science workflows into simple, chat-based natural language dialogues. MLflow, an open-source platform, has also been seamlessly integrated into the ecosystem to enable data scientists to create their own custom AI agents.  

Many Use Cases 

Kalidindi chose deviation detection as a use case for the Genedata Vico-powered SPINE Explorer. Biologics are manufactured in batches, and every stage is continuously monitored using a variety of instruments to ensure safety, purity, and consistency. 

Invariably, some of the batches fail, but AI can provide an early alert when a process is going off script, says Kalidindi. “In many cases, you can intervene and take corrective action ... to avert failures.” Moreover, that project provided a closed-loop opportunity to apply what was learned into the design of subsequent batches.   

“Based on industry numbers, we expect a good amount of ROI,” he says, with value being derived from the detection of process drifts and rapid root-cause identification analysis. In terms of time reduction, Novartis estimates saving 80% of failed batches (nine to 10 per year) and the avoidance of 14 to 15 months of cumulative trial-supply delays annually. Material cost is estimated to have been lowered by $2.9 million at each of three sites, with investigation effort savings of between 720 and 1,200 workdays annually. 

Among the other potential use cases cited by Kalidindi are optimization of chemical routes, formulation prediction, solid form selection, stability prediction, analytical methods, downstream process development, and upstream process development, as well as deviation management.   

Data scientists at Novartis are reportedly “positively energized” by what they’re able to see in the data, including data lineage and context, real-time quality and manufacturing metrics, AI-driven predictive insights, and regulatory readiness. It can also be a knowledge store of the context of conversations users have with AI agents, so even if the most brilliant scientist leaves the company, a junior staffer will be led to asking well-informed questions, Kalidindi says.