Technology
Navigating EV Manufacturing Quality with Karthick Ravichandran
Karthick Ravichandran’s core argument is one the automotive industry is learning under pressure: quality planning ...
Praveen Vasudevan is transforming life-science supply chains by applying machine learning to balance high service levels with efficiency, reducing risk in industries where stockouts can directly impact patient outcomes. His work—from optimizing safety stock with predictive analytics to advancing toward Agentic AI—shows how data-driven strategies can make supply chains more resilient, cost-effective, and patient-centric in an era of global volatility.
In the global life-sciences industry, the supply chain is more than a network of logistics; it is a lifeline. The imperative to have life-saving products available at the right place and time is absolute, as a stockout is not merely a lost sale but a potential risk to patient outcomes.
Yet, this critical mission operates under immense pressure. The COVID-19 pandemic starkly revealed the fragility of global supply chains, exposing vulnerabilities like over-dependence on limited suppliers and the risks of geographically concentrated manufacturing.
This dynamic creates a persistent tension. The need to maintain high service levels, which traditionally requires vast and expensive inventory buffers, clashes with the financial drag of carrying excess stock, especially for products with short shelf lives.
Navigating this complex landscape requires a new class of expertise that bridges deep operational knowledge with advanced analytical power. Praveen Vasudevan, a supply chain management professional with nearly two decades of experience, embodies this modern synthesis.
As an IT manager and supply chain solutions architect for a global life-science company, he leverages a robust background in both supply chain management and analytics to architect data-driven solutions. His work has consistently focused on helping US-based clients drive efficiency and tangible improvements in their supply chain processes, a mission he continues in an industry dedicated to saving lives by advancing science.
Vasudevan’s approach centers on leveraging machine-learning techniques to fundamentally reshape how inventory is managed. He focuses on analyzing the immense volumes of data generated by Enterprise Resource Planning (ERP) systems like SAP and Kinaxis, meticulously separating valuable signals from ambient noise.
By applying predictive and prescriptive analytics, he generates crucial insights into near-term and long-term supply chain performance. These insights empower business leaders to make decisions that optimize inventory, enhance working capital efficiency, and ultimately increase the probability of products reaching customers when and where they are needed.
His innovative work, which includes developing a robust IT solution for annual operating plans using Kinaxis Maestro and SAP, has earned him numerous awards and speaking engagements. This includes a first-place win at SAP ShareNet for his work on the SAP Advanced Planner and Optimizer application.
The journey toward adopting advanced analytics often begins when the limitations of established methods become undeniable. In the life-science industry, the traditional mathematics for determining safety stock have long been the standard, relying on statistical analysis of historical data to calculate the buffer needed to meet a desired service level.
While history can be a reasonable guide, its predictive power wanes in the face of unprecedented disruption. The COVID-19 pandemic served as a global wake-up call, exposing how even sophisticated algorithms based on past performance were unprepared for the seismic shifts that occurred.
Vasudevan notes that this crisis highlighted a fundamental weakness in backward-looking models. “In the life science industry, where a significant portfolio of products is used to save lives, having the right level of inventory at the right place at the right time is critical,” he explains. “While the math behind safety stock determination hasn’t changed, the science and art used have been increasingly influenced by the growing advancements in computing capabilities and the accessibility to machine-learning models.”
The key vulnerability lies in relying solely on “history.” While historical data forms the bedrock of classic safety stock formulas, it cannot account for novel events.
This realization catalyzed a shift in thinking toward actively integrating “leading indicators”—current events that signal future changes. “Let’s take the example of the COVID pandemic,” Vasudevan states. “The rising number of cases in a country today is a leading indicator that could help a life science company adjust the safety stock levels for a relevant product at a warehouse in anticipation of increased demand from hospitals and laboratories. We cannot brush these aside as outliers, especially when machine-learning algorithms can help us adjust our buffer stock based on today’s environmental and geo-political developments, such as tariffs and hurricanes.”
Improving forecast accuracy is a central goal for any supply chain organization, as it directly influences inventory levels, costs, and customer service. While many organizations continue to use time series forecasting techniques, these methods have inherent limitations, particularly the trade-off between aggregation and accuracy.
As Vasudevan explains, forecasts are almost always more accurate at higher levels of aggregation because the larger dataset makes it easier for models to identify underlying patterns. However, highly aggregated forecasts are not operationally useful for planning supply.
To address this, Vasudevan has focused on using machine learning to find the “sweet spot” in the data hierarchy where the demand signal is strongest. “Considering the product and location dimensions, manually trying to determine the right level of aggregation is near impossible,” he says. “This is where machine-learning algorithms could play a very big role. These models could redetermine the right level of aggregation across product and location using statistical parameters such as the coefficient of variation and automatically aggregate historical data to generate a forecast.”
To measure the effectiveness of these advanced models, it is crucial to use metrics that reflect their real-world utility. “One of the most useful metrics used to measure the efficacy of any forecasting model is the accuracy of the lag ‘n’ forecast – where ‘n’ represents the number of months in advance a forecast is made,” he explains. “Most organizations will try and improve their long-term forecasts to keep their supply chain stable. Tracking accuracy across increasing lags helps evaluate a model’s robustness.”
The successful integration of machine learning into core business processes depends on more than just the technical prowess of the models. A significant hurdle is the “black box” problem, where the inner workings of a complex algorithm are opaque to the business users who must rely on its output.
Building trust and driving adoption among planners is a critical challenge that Vasudevan has addressed by focusing on model explainability. Modern supply chain planning platforms like Kinaxis’ Maestro offer the ability to either use built-in ML algorithms or integrate custom models written in powerful programming languages like ‘R’.
Regardless of the model’s origin, its value is only realized if its logic is understood. “Whether these models are in-built or external to the planning platform, one of the biggest challenges is explaining the output for the users to understand,” Vasudevan notes. “This is where dimensionality reduction techniques such as Principal Component Analysis (PCA) play an important role.”
PCA transforms a large set of variables into a smaller set of “principal components” that explain most of the variance in the data, making it possible to articulate the “why” behind a model’s recommendation. This focus on explainability serves as a strategic tool for change management.
“While machine learning models such as decision trees, random forests, and others can use multiple features or attributes to classify data, PCA analyzes these attributes in the data set to identify the top few dimensions that explain most of the data,” Vasudevan says. “Once the top few features are identified, the logic behind the model outputs is comprehensible to the planners, thus driving adoption.”
The power of predictive analytics lies in its ability to answer “what might happen,” enabling organizations to manage future uncertainty proactively. However, these insights must be translated into concrete actions to generate value.
Vasudevan has achieved this by connecting predictive models to sophisticated simulation tools, transforming forecasts into tangible inventory and working capital decisions. This process begins with scenario planning, a capability that allows organizations to simulate the impact of various future events on their supply chain.
This simulation capability is fueled by a robust consensus forecast. “Demand sensing capabilities that use leading indicators to determine near-term impact on forecasts are infused into the forecasting process,” Vasudevan explains. “Demand planners incorporate insights from sales and marketing teams to build alternate demand scenarios. These are weighted and synthesized into a consensus forecast that informs safety stock and cycle stock calculations.”
Once a forecast is established, the final step is to determine the most capital-efficient way to position inventory across the network. For this, Vasudevan utilizes multi-echelon inventory optimization (MEIO), a strategic approach that manages inventory holistically.
“Multi-echelon inventory models are used to determine the optimal safety stock and cycle stock to be held in a warehouse by using not only the variability of customer forecast, but also that of the inter-site demand,” he says, “thereby optimizing the overall inventory of the product in the supply network and reducing the working capital while minimizing stock-out situations.”
The theoretical benefits of machine learning become concrete when applied to solve real-world business problems with measurable financial impact. Vasudevan highlights an engagement with a large US-based food and beverage ingredients producer that demonstrates the power of these techniques.
The company needed to optimize its cycle and safety stock to reduce costs without compromising customer service. The solution involved a bespoke machine learning model designed to bring precision to its inventory policy.
The core of the solution was the innovative use of classification algorithms to simplify the optimization problem. “One of the engagements I worked on for a large food and beverage ingredients producer in the US was on assessing the current levels of cycle and safety stock held in their supply chain and to optimize inventory levels for various service level targets,” Vasudevan recalls. “The solution developed was a machine learning model that used ‘decision trees and random forests’ and ‘k-nearest neighbors (KNN)’ supervised classification techniques to classify products into aggregation groups.”
This approach first grouped thousands of individual products into rational families based on shared demand characteristics like volume and variability. This pre-processing step was critical, as it created a more manageable set of product groups with stronger aggregate data signals.
“These models grouped products by shared demand characteristics, enabling more precise calculation of annual demand and safety stock levels,” Vasudevan explains. “The output was then used by the supply chain managers to select the candidate parts where the inventory levels were adjusted. The projected working capital reduction was estimated at two million US dollars annually.”
In a world of constant change, a machine learning model is only as good as its ability to adapt. A static model will quickly become obsolete as market conditions evolve.
Ensuring these analytical systems remain resilient requires a two-pronged approach: continuously feeding the models with relevant, real-time data and empowering human planners with the tools to manage by exception. The first pillar of adaptability is the integration of leading indicator data that goes beyond historical sales, allowing models to sense and respond to disruptions as they happen.
“Integrating leading indicator data that impacts demand via machine learning models ensures that the models adjust near-term customer forecasts,” Vasudevan states. “Supply chain planners can use this information to adjust production volumes or redirect inventory to the correct regional warehouses to reduce excess and obsolescence risk.”
The second pillar involves creating a symbiotic relationship between the AI and the human expert. It is neither feasible nor desirable for planners to manually review every model output.
Instead, their expertise should be focused where it is most valuable. “It is important to ensure that the model outputs are explainable,” he says. “It is equally important that planners are equipped with exception-based reports—dashboards that highlight only the most critical deviations—allowing them to focus on high-impact decisions.” This strategy optimizes both computational and human resources, elevating the planner from a data processor to a strategic problem-solver.
The life-science supply chain operates under a unique tension. On one side, products are often highly specialized with few substitutes, making availability a non-negotiable priority to protect patient health and safety.
This imperative encourages high safety stock levels. On the other side, many of these products have short shelf lives, meaning large inventory buffers create significant financial risk from potential expiration and obsolescence.
Vasudevan identifies this paradox as the central challenge for the industry. “Unlike industries such as FMCG, where consumers have many choices, the life-science industry has a bigger impact on people’s lives, where the products and offerings are unique and seldom interchangeable,” he observes. “As a result, organizations in this industry tend to prioritize high service levels, resulting in higher safety stock. This naturally leads to inefficiencies in the supply chain, as the risk of stock-out is a lot higher than carrying excess inventory.”
In this high-stakes context, improving forecast accuracy becomes the primary tool for risk mitigation. By reducing the fundamental uncertainty around future demand, machine learning addresses both sides of the risk equation.
“One of the important ways to reduce the inefficiencies in the supply chain is by improving forecast accuracy,” Vasudevan states. “With more accurate forecasts, organizations can carry significantly lower safety stock levels without compromising availability—thereby reducing working capital and minimizing obsolescence.”
The evolution of artificial intelligence in the supply chain is accelerating, moving beyond prediction and into the realm of autonomous action. While machine learning has been game-changing for forecasting, the rise of generative AI has made interacting with complex data more intuitive.
Vasudevan sees this as a stepping stone to the next frontier: Agentic AI. Unlike generative AI, which responds to prompts, Agentic AI systems are designed to be goal-oriented, capable of planning and executing multi-step tasks autonomously.
This represents a significant leap in capability. “Rapid advancements are being made in the field of Agentic AI as well, where, in addition to generating insights, the machine learning models can execute decisions within certain limits,” Vasudevan explains. “This does not mean that Agentic AI will replace humans in the supply chain. Instead, they will equip humans with the ability to process and make sense of large volumes of data to make better decisions.”
This technological progression signals a fundamental shift in the role of the supply chain professional. As AI takes on more of the tactical execution, the value of human experts will migrate toward higher-order, uniquely human capabilities like strategic direction and ethical oversight.
“In the life science industry, machine learning models and AI can play a very important role in improving product quality and speed to market of new products,” Vasudevan says. “With access to generative AI, supply chain professionals can move on from understanding syntax to understanding data. Time to insights has drastically reduced, leaving more opportunity to derive meaningful insights and make better decisions.”
In the face of unprecedented volatility, legacy supply chain strategies are proving insufficient, particularly in the critical life-science sector. The work of experts like Vasudevan demonstrates a clear path forward: harnessing the predictive and autonomous power of machine learning to transform supply chains from reactive functions into proactive, resilient, and data-driven strategic assets.
This evolution allows organizations to navigate the complex trade-offs between service, cost, and risk with newfound precision. Ultimately, this transformation is achieved not by replacing human expertise but by augmenting it, paving the way for a future where technology and human ingenuity combine to deliver life-saving products to patients more efficiently and reliably than ever before.
