“The blind spot isn’t that the risk exists; it’s that the tools most networks plan with are too simple to reveal it before it becomes real.”
Dr. Narendiran Sivanesan is CEO and co-founder of tulanā, a decision intelligence startup that fuses machine learning, mathematical modeling, and AI to help supply chain and logistics companies make robust, cost-effective decisions under uncertainty, including capacity planning and inventory optimization.
He was previously a lead data scientist at BCG Gamma in NYC, the AI and machine learning division of Boston Consulting Group, and brings over ten years of experience applying AI to build enterprise applications. Narendiran holds a PhD in mathematics from Friedrich-Alexander-Universität Erlangen-Nürnberg, Germany, and an MBA from Collège des Ingénieurs in Paris, France.
Ahead of his session at CDMO Live Americas 2026 in Boston, Narendiran spoke with PharmaSource about the rising cost of getting capacity decisions wrong, why large language models are the wrong tool for planning problems, and how his team helped one biopharma client uncover risk its own planning process had missed.
The Stakes on Capacity Decisions Have Gone Up
Three things have changed in pharma capacity planning over the past five years, Narendiran says, and they compound each other. First, the cost of a wrong call has risen sharply. Tariffs and reshoring have driven roughly $480 billion in new US manufacturing investment since early 2025, on top of GLP-1-driven builds like Eli Lilly’s $27 billion and Novo Nordisk’s $9 billion.
“A new facility takes five to ten years to build and can’t be resized easily once it’s running. So today’s capacity decision is a decade-long bet, and the bets are much bigger now.”
Second, interest in AI-driven planning tools is a lot higher than actual adoption. A recent survey of pharma and biotech tech executives found 57% expect strong returns from AI and analytics on demand forecasting this year, but only 24% report measurable value so far. Third, reliance on CDMOs has grown alongside both trends, turning capacity decisions from a simple build-or-don’t-build call into an ongoing negotiation over how much to commit in-house versus flex externally.
Narendiran points to a rare-disease biopharma client as the clearest illustration. Checking their production plan against a single demand scenario took the client’s team multiple days by hand, so they could only test a handful. tulanā ran all 243 relevant scenarios automatically. One showed the client’s need for external manufacturing capacity growing from 26% to 64% of output within five years, against a two-year CDMO tech-transfer lead time. “That’s a cliff you need to see three or four years out, not when it’s already happening,” he says.
What Large Language Models Are For, and What They’re Not
Narendiran is careful to separate what people mean by “AI” from what he thinks is actually needed for these decisions. Large language models, he argues, are the wrong architecture for capacity planning.
“LLMs are pattern-matchers — good at language, not at guaranteeing a plan which respects your actual constraints. Capacity planning is a combinatorial optimization problem: thousands of interdependent decisions on production volumes, timing, and site allocation, all bound by hard limits like plant capacity, contractual minimums, and shelf life. An LLM can describe a plan that sounds reasonable while quietly violating several of those limits, because nothing forces it to respect them.”
That’s not to say LLMs have no place here,” he adds. “They’re useful for the opposite end of the problem: helping a planner formulate a scenario or a constraint in plain language, so you can set up and iterate on a question faster. The model still has to solve it. The LLM just helps you ask it well.
tulanā’s approach, he says, never allows a large language model to touch the plan directly. Instead, the team works with a client’s own team to build a mathematical model of the actual shop floor and network, encoding real-world constraints into a “strongly typed system.” These constraints become business rules that the code cannot violate and are checked before anything runs.
“The LLM never touches the plan itself. It facilitates interaction with the mathematical model: helping translate a planner’s question into the model’s terms, and helping interpret what comes back. The model is the one that decides, and it can only output something that satisfies every constraint, the same way a compiler won’t build code that doesn’t type-check. We want AI to speed up the analysis, not to quietly pick your production plan.”
The upside of leaning on math rather than a language model, he adds, is that the resulting plan is fully auditable. The system deals with probability and uncertainty going in, but the decisions that come out are deterministic and traceable.
The Industry’s AI Alarm
The interview came in the middle of an unusually candid fortnight for AI safety warnings. In the week before speaking with PharmaSource, several senior AI researchers said publicly they believe frontier AI could pose an existential risk within the next decade, and Anthropic’s chief executive called for the industry to deliberately slow the pace of capability development. Anthropic also published a threat intelligence report documenting attempts to misuse its models for biological weapons research. A stark illustration of what the industry calls the “dual-use” problem, where the same knowledge needed to cure a disease can be used to engineer one.
Narendiran, who has several friends working across Silicon Valley AI labs, said the concern is genuine and that he shares some of it.
“The dangers are multiple-fold. The obvious one is that someone can build a bioweapon — bad actors. A second one is a little more futuristic: we’ve already seen examples where AI agents have escaped their own sandbox, hacked their way out of a controlled environment. That’s the Terminator scenario, which I think is very futuristic but not completely implausible. And the one I think is the scariest and the most probable is the amplification of fake news and civil unrest by AI agents that post independently online and amplify messages people can’t tell apart from the real thing, because they sound so human. We’re already a polarized society, and this will make it worse.”
That same quality, sounding convincingly human, is also what makes large language models unsuited to capacity planning, he argues, tying the point back to his core thesis: LLMs are optimized to produce plausible, human-sounding language, not to guarantee that a decision respects hard constraints.
On the dual-use question specifically, Narendiran draws a line between what AI safety researchers call “alignment” and what his own field calls “formal verification.”
“What we’re doing is making sure that code and planning systems produce correct outputs, outputs that respect your constraints. In the LLM case, it’s more about guardrails, or what AI researchers call alignment: making sure the model’s behavior lines up with human morals and ethics. These models are, at their core, optimizing for a reward. If you tell one to produce as much of something as possible, it will find ways to do that you never intended, unless you’ve been very careful about the guardrails. Formal verification goes down a different path. The LLM formulates and proposes; the math vindicates it, ruling out anything that violates your constraints.”
It’s a distinction he returns to when asked to place tulanā’s own technology against the backdrop of those warnings.
“We are working with mathematical models. You have to read through the math — I know, not most people’s favorite pastime activity — but on the upside, it makes the decisions auditable. We deal with probabilistic things, so we can account for uncertainty, but the decisions themselves are deterministic, and you can actually audit them. I think that’s the right kind of technology for these kinds of planning questions.”
The Billion-Dollar Blind Spot
The case study behind Narendiran’s CDMO Live Americas session title came from that same rare-disease biopharma client, who had built a ten-year capacity plan in Excel and believed it held up, particularly in lower-demand scenarios.
“We took their actual plan and loaded it into our mathematical model to check it against their real constraints. The scale of that model tells you why Excel couldn’t catch what it missed: 243 demand scenarios across their product portfolio, roughly 1.5 million scheduling decision variables, and close to 1.2 million constraints representing their actual manufacturing network. A spreadsheet can hold a handful of scenarios and a few dozen constraints at most; it was never going to surface violations buried in a system that size.”
It didn’t. Even in the base case, with no additional demand upside, the client’s plan was already violating safety stock minimums and running utilization at one site at 95%, against their own 80% target ceiling. “That’s not a stress scenario,” Narendiran says. “That’s the plan they thought was safe.”
Across the full range of demand scenarios, profit at risk from unmanaged capacity gaps came out between roughly $600 million and $15 billion over the ten-year horizon (profit, not revenue), depending on how demand for the client’s drug in trial actually plays out.
Putting a Number on Tariff Risk
Narendiran applies the same scenario-based approach to geopolitical uncertainty. Rather than forecasting a single tariff rate, tulanā runs thousands of Monte Carlo draws across tariff timing, rate, cost, and construction delays, then calculates the five-year net value a business ends up with in each simulated future.
“Downside risk isn’t the average outcome; it’s the worst plausible one, and how much it still nets you.”
In an illustrative case study, tulanā compares two strategies for a mid-sized pharma company sourcing API from China: keep the current network and absorb tariff costs, or dual-source through a second CDMO in India. Even in its worst simulated outcome, staying with the status quo still nets $440 million in value over five years. Dual-sourcing’s worst outcome nets $980 million, more than double, because it removes most of the tariff exposure rather than absorbing it.
“The two ranges of outcomes don’t even overlap: the worst case for dual-sourcing beats the average, expected case for doing nothing. That’s what makes it a real number instead of a gut feeling. You can compare exactly how much each choice is worth even when things go wrong, instead of debating opinions about tariff policy.”
What Actually Changed for the Client
Three things, according to Narendiran. First, the client got a plan they could trust, after the model exposed the safety stock and utilization violations their Excel-based plan had missed. Second, their approach to the question itself changed.
“Before, capacity planning meant building one plan and hoping it held. After, it meant checking every plan against a full range of possible futures before committing capital. The client described the work as foundational to how they now think about these kinds of tools.”
Third, and this is what Narendiran says drives investment decisions, the model doesn’t just show where capacity risk sits; it shows when. With a two-year CDMO tech-transfer lead time, knowing that external capacity needs would grow sharply within five years wasn’t just a risk flag; it was a trigger, telling the client the moment to start that CDMO conversation or option a shell facility rather than finding out too late to act.
“That’s the difference between a warning and a decision.”
Capacity planning & network optimization
Optimize your entire value chain with proactive resilience.
Plan your entire supply network over 5 to 10 year horizons. From API synthesis through formulation, packaging, and distribution, while accounting for demand uncertainty, new drug launches, regulatory approvals, and capacity investments.
Supply planning & production scheduling
Maximize throughput while maintaining compliance.
Determine optimal production quantities, batch sizes, and scheduling sequences across multiple facilities. Account for setup times, changeover costs, equipment qualifications, and cGMP requirements.
Inventory optimization & safety stock planning
Reduce waste from expiry while ensuring availability.
Optimize safety stock levels, reorder points, and inventory positioning across your network. Balance service level targets against carrying costs while respecting shelf-life constraints and cold chain requirements.
