Product and tech teams can significantly enhance AI's ROI by building on small language models—if they can create the high-quality training data these models require.
How Boosted.ai launched a better, faster, AI investment assistant and cut costs by 90% at the same time.
The name of the game at this point is ‘how many new products can you release?’ Everything that allows you to move quicker is a win.
Jonathan, what problem were you trying to solve?
We had a successful AI portfolio assistant that analyzed 150,000 data sources and 60,000 stocks overnight, cutting research time for asset managers from a week to a day. Demand was growing for broader coverage and faster insights, but scaling the models required GPUs at an impossible scale. We needed a way to get through the last mile of data quality. Our SLM was performing well, but without trusted human evaluation, we couldn’t push accuracy high enough to support real investment decisions.
Why did you choose Invisible?
Crowd platforms and in-house teams couldn’t deliver the scale or sophistication of data we needed. Fortunately, AWS connected us with Invisible, who gave us domain-specific finance experts plus metadata-rich explanations that taught our own national language processing team new ways to improve the model.
Our NLP leads hadn’t realized something, but there was an explanation in the metadata that they read and said, ‘Huh! I could get behind that.’ So there were things that even they learned… Explaining the explanations, I think, really helped.
What was the commercial impact?
Our customers get faster, evidence-based market signals, and we were able to scale this reliably at 90% lower cost.
With Invisible managing data quality, our product and engineering teams could focus on ensuring they had the smallest possible model, and the highest performing AI Investment Assistant.
