SigOpt

SigOpt was built around a simple observation: companies building complex AI and machine learning pipelines often fail to get peak performance because of the many knobs and levers that have to be tuned. Scott Clark, a PhD in applied mathematics who had spent years applying optimal learning technologies in industry and academia, saw that human experts were poor at tuning these systems. As Scott put it, "humans are pretty bad at doing 10 dimensional optimization in their head."

The product is optimization as a service, delivered through an API. Rather than ingesting raw data and replacing a client's existing model, SigOpt uses a black box approach. It only looks at inputs and outputs of a system and applies an ensemble of Bayesian optimization strategies to find better configurations. Clients pay a subscription based on the number of models they tune each week. Scott explained that the company's domain expertise lies in the optimization layer itself, which allows it to work across industries without deep assumptions about the underlying model.

SigOpt focused on top-down enterprise sales. The team engaged executives excited about artificial intelligence, then moved to the practitioners who could bolt the product onto their systems. Early and current customers included Prudential, MIT, and Hotwire, with Huawei also named, along with credit card companies and algorithmic trading firms. Scott said the company had about a dozen paying customers and that enterprise plans typically started around $10,000 per month. Three or four salespeople worked to close deals, and the team visited customers in person in cities like New York and Houston.

The company reported no churn so far. Scott argued that once customers embed SigOpt in their systems, they are unlikely to return to manual brute-force tuning, which makes the product sticky. Annual contracts were preferred and incentivized, often with a discount. Headcount was the largest expense, with the team of 13 mostly based in San Francisco and one remote member in Spain. The host's rough math suggested headcount alone could run around $130,000 per month, which Scott did not dispute.

At the time of the interview, SigOpt had raised about $8.8 million, had gone through Y Combinator, and had its Series A led by Andreessen Horowitz the previous July. The company was not actively raising a Series B and described itself as building a sustainable business. The host estimated monthly recurring revenue at around $50,000.