How Amplified.ai Keeps Innovation Out of the Shadows

Customer

Amplified.ai

Founders

Chris Grainger

Yasuyuki Oikawa

Samuel Davis

What's the most important part of our patent system? If you said: "protecting and sharing good ideas," Chris Grainger agrees with you. The founder of Amplified uses AI to make it easy to search patents that have been granted by the USPTO.

Searching for Patents Is Hard

Why do we need a tool like Amplified? Doesn't the patent database allow for basic things like…keyword search? It turns out that the U.S. patent system isn't as easy to navigate as you might think. While there's an obvious reason to file a patent—protection for your ideas, and a monopoly on sales that use them—once you've got that legal protection, there's also an incentive to hide it away in your sock drawer. Patent holders want to keep their trade secrets close and prevent competitors from using their work as a launchpad, because it only takes a small-ish tweak to a patented invention for it to become fair game for anyone to sell (or patent!) themselves.

Amplified Finds Obscure Patent Applications

The increasing ease of information retrieval over the last 50 years has pushed patent holders (and, let's be real, their attorneys) to get pretty tricky in how they register their patents. It's now common practice to file under a zillion derivations of a company name, or use tortured language to describe the underlying invention, all in the service of making keyword search useless. That's where Amplified comes to the rescue: they use their own bespoke natural language model to crack this behavior. Using their platform, you can just search "hovercrafts" or "real-time language translation ear piece" and bypass any shenanigans to learn the real state of the patent landscape. This is a game-changer for inventors and anyone else who works in R&D.

Small Team, Big Data

Amplified initially grew out of Grainger's PhD work in 2013, when he used patent data for econometric models. It's since grown into a knowledge platform that helps its users sift through around 150 million registered independent patents. Amplified's small engineering team (two brave souls, including Grainger) had to find creative ways to help their customers wade through this Olympic-sized data pool. The key? Amplified relies on vector embeddings to pull together any inventions that have similar technology (even when the language of their patents doesn't indicate the link).

Custom LLMs on Fly.io GPUs

Amplified uses their own bespoke LLM, rather than ChatGPT or similar models, since they've found that newer generative models aren't fast or powerful enough for the very large scale embedding models they run. They serve it out of an Elastic Search cluster, and then use that data to run other models like classification, clustering, and more.

Self-Scaling Infrastructure Without an Ops Team

Amplified's been running their stuff on Fly.io for over a year. As a (self-described) small team with limited resources, having everything in a monolith on Fly.io makes things significantly easier for Grainger and his team to operate. Says Grainger, "Fly.io is really good at deploying apps so we don't have to be, and has made smart default choices about multi-region, load-balancing, and networking." Their core infrastructure makes heavy use of the FLAME pattern in Elixir, which helps them scale elastically as needed without the pain of managing endless microservices. For example, Amplified's embedding workflow uses Oban to run tasks in batches as needed, spinning up a swarm of Fly.io GPUs as needed, then back down to zero when they're not. They've run over 80 million documents (many of them the size of Moby Dick) through this hyper-efficient embedding process without breaking a sweat.

7.5 Terabytes of Data Stored on Tigris

Another way they handle their massive data load is with Tigris Data, which Amplified uses to solve and scale their document storage. Amplified stores documents in a massive Postgres data warehouse. But it's no small potatoes; this is a 7.5 terabyte single node instance, and getting large amounts of data out of it rapidly was, as Grainger put it, "the bane of my existence for years." Amplified doesn't need this massive Postgres instance running constantly, so they store each document as an XML file on Tigris, opening up essentially unlimited horizontal scaling.

Plenty of Room to Scale

The "innovation industry" is only picking up speed; those 150 million patents represent a 5-fold increase in the last 50 years. Amplified's platform helps inventors and R&D departments keep up with the real pace of technology…which means our hovercars probably aren't that far off.