Cambrium uses genetic engineering and computational models to design proteins that can become physical materials, starting with vegan human collagen for cosmetics.
2
The engineering constraints in biotech come from expensive experiments and valuable data, rather than high request volume or low-latency model serving.
3
Protein sequences have a language-like structure, which allows generative models to propose proteins with specified properties that can then be tested in the lab.
Summary
Pierre Salvy describes Cambrium's work at the intersection of biotechnology, engineering, and AI. The company designs proteins that can be produced by genetically modified yeast and used as physical materials. Its first product, Nova Co, is vegan human collagen for cosmetics. Salvy explains why the company's platform has different needs from a typical ML platform. A failed pipeline can lose experimental data, while a failed prediction usually costs money and time rather than putting lives at risk. Cambrium uses robots and structured records to make lab work more data-driven. Salvy also compares proteins with natural language because both can be represented as sequences of letters with an underlying grammar. Models can now generate protein sequences with target properties, although lab experiments are still needed. He connects the work to future applications such as sustainable plastics, adhesives, food products, and materials for space, while describing regulation, cost, and time to market as major limits.
Salvy moved from applied mathematics into industrial biotechnology
Pierre Salvy studied applied mathematics and applied physics at the École des Mines de Paris, covering areas such as materials science, nuclear physics, biotechnology, and economics. During his master's studies, he worked on a project that produced jet fuel from genetically modified yeast. The process used the same basic idea as brewing beer, except the yeast was engineered to make jet fuel instead. Salvy built mathematical models of the process. He later completed a PhD at EPFL and joined Cambrium in Berlin in 2020, shortly before defending his thesis. He became the company's second employee and initially led computational biology.
His move into management required learning how people work
Salvy says the largest change from being an individual technical contributor to leading engineering was learning to manage people. He quotes Cambrium co-founder Ruben reminding him of a time when he said, "I study computers, not people. I don't know how to handle that." Salvy learned that management requires understanding how people react in situations he has not faced before. For difficult conversations, he tries to identify what each person does not understand about the other's reasoning. He treats disagreement as an analytical problem where both sides explain their assumptions. This approach has helped during stressful startup decisions and mistakes.
He stays technically current through limited hands-on work and trust in specialists
Salvy describes the difficulty of staying current after leaving academic research. He still spends time interacting with new technology, including LLMs, LangChain, and different model backends. That hands-on work helps him understand what systems can and cannot do, including failure modes such as ChatGPT trying to convince him that it was alive. He does not believe he needs to use every tool every day. His engineering team stays closer to the technical frontier than he can, so he relies on their expertise while contributing through prototypes, direction, and management.
Cambrium uses LLMs for research assistance and protein generation
Salvy says Cambrium's LLM work began with a disagreement between himself and Charlie, the chief scientific officer. The discussion led to several practical use cases. One system combines a LangChain stack with company information from experimental data, Slack messages, and Notion pages. Researchers can ask questions about past work, retrieve protein sequences, summarize results, or draft a LinkedIn post from internal information. Salvy describes this as a useful search engine rather than the company's main route to changing material production. Other projects include an LLM-based research assistant and machine learning systems that generate proteins.
The first commercial product is vegan human collagen made by engineered yeast
Cambrium incorporated in September 2020 and commercialized its first material biotechnology product in less than two years. Nova Co is vegan human collagen for cosmetics. Collagen is the most abundant protein in the human body by weight and is associated with skin structure. Cambrium takes a collagen sequence, inserts it into yeast, feeds the yeast sugar, and extracts and purifies the resulting collagen. Suppliers then use it in cosmetic products. Salvy presents genetic engineering as a practical way to produce animal proteins without using dead animals, alongside examples such as insulin and enzymes used in washing powder.
Regulation shapes which markets a biomaterials company can enter first
Salvy says Cambrium began with cosmetics because the category allows faster product cycles than medical products while still having room for innovation. Medical products can take five to eight years to reach the market, which is difficult for a startup to fund in Europe. He also describes regulatory barriers for food biotechnology. Companies such as Formo can reproduce cheese molecule by molecule without animals, yet Salvy says they have to commercialize outside Europe first. Cambrium's longer-term path is to move from specialty, higher-price materials toward lower-price commodity materials as production costs fall.
The main engineering problem is preserving expensive experimental data
Cambrium does not need infrastructure designed to serve millions of users at low latency. Its users are scientists, and a service outage usually means they ask the engineering team to restore it rather than switching providers. The expensive part is the biological experiment. DNA, robots, consumables, lab space, and equipment all cost money. Salvy says every data point must be preserved because a failed pipeline can lose information that cannot be recovered. The company works with machine outputs such as CSV files and with scientists' experimental notes. Robots can record settings and actions that scientists would otherwise need to enter manually.
Protein sequences have a grammar that generative models can learn
Salvy compares proteins with natural language. Proteins are sequences made from 20 amino acids, which researchers commonly represent with letters. Random letters rarely form a useful sentence, and random amino acids usually produce a nonfunctional structure. Protein engineering tries to learn the sequence rules that lead to functions such as hair flexibility, skin elasticity, or oxygen transport in blood. AlphaFold helped predict a protein's shape from its sequence. Generative protein models go further by proposing sequences with particular properties. Cambrium can express those properties as mathematical loss functions and use models to generate candidate proteins for laboratory testing.
"You have to interact with the technology to keep your touch, but you don't have to be playing with it every day to stay at the bleeding edge."Pierre Salvy20:04
Who should watch
You are building ML systems for scientific or laboratory users and need to understand why data capture and experiment tracking can matter more than serving throughput.
Your team is considering generative models for biology and wants a concrete description of how generated protein sequences connect to lab experiments.
You work on sustainable materials or biotech products and want to hear how regulation, capital costs, and market choice affect the route from research to commercialization.
