Challenge Provider: Umami Bioworks (Shivansh Singhal)
Team: Ivor Spector, Alex Yu
Mentors: Charlotte Capelle, Tatiana Codreanu
The Challenge
Scaling cultivated seafood remains constrained by the high cost and inefficiency of optimising growth media and bioprocess conditions through trial and error. Media ingredients can account for up to 80% of production costs, while limited data availability makes it difficult to balance yield, cost, and flavour.
The challenge aimed to develop an AI-driven modelling approach that could streamline optimisation, reduce experimental burden, and accelerate the transition toward sustainable, affordable cultivated seafood.
The Approach
The team designed a prototype AI platform, PilotPath, envisioned as a digital twin for cultivated seafood bioprocessing. It integrates limited yet diverse data sources, including historical media formulations, ingredient pricing, bioreactor time-series data (cell growth, viability, spent media composition), and sensory and chemical flavour profiles. Using predictive modelling, the platform identifies high-impact media components and correlates them with growth and sensory performance.
The model was refined iteratively using simulated client-specific datasets, ensuring data privacy while allowing customised recommendations. An embedded optimisation engine then proposed media formulations balancing yield, cost, and flavour, guiding users toward the next most informative experiment to run.
The Outcome
The resulting PilotPath prototype demonstrated how AI-guided experimental design can dramatically reduce time, cost, and resource use in cultivated seafood R&D. By providing actionable insights and targeted experiment suggestions, the system shortens the optimisation cycle from months to weeks.
It represents a first step toward autonomous bioprocess development in the cultivated protein space, helping startups overcome data scarcity while advancing data-driven, scalable, and flavour-optimised seafood production.
Impact & Next Steps
This project tackles the common challenge of sparse, fragmented datasets in bioprocessing. Instead of relying on fully comprehensive models, which are rarely feasible for early-stage startups, it uses adaptive, stepwise models to identify the next best experiment. This approach reduces the need for costly trial runs, supports decision-making under uncertainty, and improves process performance with each iteration.
Next steps include securing additional datasets to strengthen the initial model, developing a user-friendly interface, and executing the first guided experimental trial.