What I've built, and what it took to build it.

Process characterization (PC) studies prove that a biologics manufacturing process is robust before it goes to commercial scale. They're also expensive: a standard design can run 75 to 90 bioreactor experiments per study, and Amgen's biologics pipeline was growing faster than lab and workforce capacity could scale to match. The question behind my thesis was simple to state and hard to answer: could a model that already understands some of the underlying biology need meaningfully less data than one that starts from nothing?

I built hybrid mechanistic and machine learning models using AMGEN's DataHowLab (DHL) platform, which combines a mechanistic simulation backbone with a learned correction layer trained on time-series bioreactor data (pH, temperature, VCD, and other conditions) rather than idealized setpoints. To run this at scale I wrote a modular Python framework: a dedicated API layer that abstracted DHL's client for project discovery, dataset lookup, and model execution, separating low-level platform interactions from the higher-level experimental logic so I could sweep training-set sizes and sampling strategies programmatically instead of clicking through a GUI hundreds of times. The team worked in an agile cadence, three standups a week plus ad hoc pairing sessions, and I held every function I wrote to full unit test coverage, which mattered more than usual given how easily silent data errors can propagate through a modeling pipeline.

On the experimental design side, I benchmarked the hybrid models against Amgen's standard JMP regression approach across eight product quality attributes (PQAs), compared multiple training-subset selection strategies (D-optimal, Sobol, random, and exterior-plus-D-optimal designs), and tested whether augmenting a small set of PC experiments with 24 upstream commercial process development (CPD) runs as a prior improved sample efficiency. I also ran a training-data sufficiency study to find the point of diminishing returns.

Models trained on roughly 35 PC experiments, augmented with the 24 CPD experiments as a prior, matched or exceeded JMP accuracy for the best-performing PQAs, a 54 percent reduction relative to the 75-run design analyzed and a 61 percent reduction relative to a standard 90-run campaign. Learning curves flattened around 35 to 45 experiments on average, though convergence varied by attribute. A few findings surprised me: measured time-series inputs consistently outperformed idealized setpoint perturbations, space-filling sampling strategies beat boundary-focused designs, and a shared mechanistic backbone across all eight PQAs outperformed attribute-specific tuning, meaning the architecture mattered more than the fine-tuning.

Beyond the accuracy numbers, I built a transparent, parameterized economic model to translate experimental reduction into operational terms: fewer runs per study compresses execution timelines, accelerates CMC milestones, and frees bioreactor and labor capacity to absorb additional studies without proportional headcount growth. I also designed the change-management and adoption roadmap for rolling the tooling out across Amgen's global network, aligning scientists, managers, and quality and regulatory stakeholders on a phased implementation plan, since a model nobody trusts or knows when to use doesn't save anyone anything.

Most of my clients were healthcare and life sciences organizations trying to translate scientific or strategic promise into something operationally real: a commercial launch plan, a new business unit, an operating model that could actually run. The gap between a good strategy slide and an organization that can execute it is where most of my four years there were spent.

The work spanned healthcare strategy, digital health platform design, and cell and gene therapy commercialization, with a few adjacent engagements in supply chain risk and technology due diligence along the way. A few representative projects:

Alongside client work, I helped grow Deloitte's Cell and Gene Therapy practice from $100M to $250M in annual revenue, coordinating strategy with 35 practice leaders and leading a project team of six through executive strategy sessions. I also started Deloitte's Synthetic Biology Food-Technology Industry Working Group from scratch, which grew to 30-plus members across 15 organizations, and ran a monthly onboarding series that helped new hires build community, something I wish had existed when I started.

After three years at Deloitte I wanted to go deeper technically without giving up the strategy and operating instincts I'd built. LGO is a dual degree built specifically for that trade: an MBA from Sloan and a Master's in Chemical Engineering, earned together over two years rather than sequentially, anchored by a six-month internship embedded inside a partner company's actual operations instead of a summer project on the side. That structure, engineering and management trained in parallel and stress-tested against a real operating environment, is what made it the right program instead of a standard one-year MBA.

Coursework spanned both halves of the degree: Advances in Biomanufacturing, Design Principles in Mammalian Systems and Synthetic Biology, and Chemical Reaction Engineering on the technical side; Competitive Strategy, Global Supply Chain Management, and Operations Strategy on the management side; plus courses built to sit at the intersection, like AI and Machine Learning in Molecular Engineering and Cell Biology. The LGO cohort itself, around 45 people, was almost entirely engineers and scientists who had also spent time in operating or strategy roles, which made most classroom discussions feel closer to a working session than a lecture.

I chaired the LGO Seminar, sat on the LGO Internship Committee, and, somewhat unexpectedly, was elected president of the Sloan Golf Club, which mostly meant continuing to organize events I'd have wanted to attend anyway. The six-month internship requirement is where the manufacturing and operations training got tested directly: mine was the process characterization work at Amgen, detailed above, which became both my thesis and a live business case.