The Scene
Generative AI startup Sweetspot, a search engine that uses large language models to look for specific U.S. government contracts, has simplified a once tedious task.
And while Sweetspot is aimed at businesses that bid on contracts, which is a $650 billion business, it could also be useful for lawmakers, journalists, and others that are keeping an eye on where federal tax dollars are going.
But none of that is why I was interested in Sweetspot. At a very high level, it’s an example of the wave of change that’s about to come. Large language models are becoming a new layer between people and technology that will simplify everything we do on the internet and with computers.
Below is an edited conversation with Sweetspot’s co-founders, Sachin Subramanian and Philip Kung.
The View From Sweetspot
Reed Albergotti: Tech companies are increasingly looking for government contracts and there are geopolitical reasons for that. What’s a real world example?
Subramanian: The government is trying to install a lot of facial recognition software of late. The contracts on Sam.gov are titled with very obscure names. Sometimes people just say, ‘software needed to detect x and y.’ But with our software, if you just type in ‘facial recognition,’ it matches businesses with all those relevant contracts.
Sometimes the government has some data that they need analyzed. Businesses have a tough time finding these contracts because there are so many things they could type in and it takes a while to find all these relevant contracts.
Q: How does it work? Do you do a bunch of training on the language used in government contracts or can you plug in existing models like GPT?
Subramanian: We pre-process the publicly available government data in our unique way, such that it allows search terms to get easily filtered out by our AI. So the AI is able to understand this data better.
Q: What does pre-process mean?
Kung: Our secret sauce is being able to determine what’s relevant, what’s not relevant in a contract. Roughly 80% of it is things like general amendments or best practices, but it has nothing to do with what the government wants. We’re able to extract that relevant 20% and that reduces a lot of noise.
Q: And you’re using AI to do that?
Kung: Yes. Our AI understands what’s important and what’s not important. Then our second step is we have a way to use AI to search across millions of indexed documents quickly to be able to find the relevant contracts quickly.
Q: And are you using any outside services for this?
Kung: We do use GPT for some scenarios, but for our core product, we fine tune our own models.
Q: Do you need experts in government contracts to train it?
Kung: No, we fine tune models that generate numerical representations of the data, known as “embeddings.” There are models out of the box that create these embeddings pretty well. Over time, you instruct the model, ‘these are the bad ones that I wish you didn’t recommend me; these are the good ones, I wish you’d recommend to me more.’ And over time, if you build that data more and more, and you feed it back to the model, it’s able to understand more precisely what you are looking for in a search dataset.
Our next step is bid proposals, where we’ll use an LLM to write the bid proposal for a contract. That’s when we’ll have to rely a lot more on the seasoned veterans who understand what makes a good contract proposal.
Q: How much does it cost?
Subramanian: $400 a month or $4,200 a year (though it’s offering discounts now).
Q: Is it paying off for customers so that the monthly fee pays for itself?
Subramanian: It’s too early to talk about it. We’ve been in production for one week. (We spoke two weeks ago)
Q: What made you want to do this?
Subramanian: We had a friend who worked at a real estate firm in LA and his job was to bid on commercial real estate contracts that were put out by the government. He explained how much of a hassle this was. You couldn’t just go on Sam.gov and magically type in commercial real estate in Los Angeles and expect to get any relevant results.
Q: So is that the target customer?
Kung: I think it really helps with small to medium-sized businesses. There’s so many contracts being uploaded to Sam.gov, from real estate to medical care to different parts for automobiles. And those are the kinds of things that are hard to find.
This is a great example. We typed in Orange Is the New Black into our search engine. It came back with federal prison contracts, specifically for women’s prisons. It shows that the AI is not looking for specific keywords because it understands that Orange Is the New Black is a show about women’s prisons.