Ontra Ask Luke Lefebure his favorite application of machine learning, and he will inspire you with what’s possible using today’s generative artificial intelligence technology and large language models. His passion for the power of machine learning directly impacts the lawyers and contract negotiations at the heart of Ontra’s business. Find out how he and the Ontra ML team take advantage of the latest tech advancements to solve complex problems.
You have vast experience as a machine learning engineer, both at Ontra and previous companies. What opportunities have you enjoyed as a machine learning engineer at Ontra, and how have you built upon your previous machine learning engineering experiences?
For the majority of my career, I’ve been a machine learning engineer focused on the sub-field of natural language processing. That’s machine learning applied to text, such as legal contracts, news articles, and government reports – there are so many applications across industries.
When I joined Ontra, the machine learning team was expanding to embed across the business. We are exploring opportunities to help areas of the business that haven’t benefited from machine learning to date.
I’m embedded in our negotiations team under Contract Automation, and our team’s job is to accelerate the negotiation of routine contracts. In the past, we applied machine learning at the end of contract negotiations by summarizing them. Now we have a broader mandate to make improvements across the entire negotiation process. With that in mind, we introduced a feature called “precedent identification,” which suggests similar contracts to lawyers at the start of negotiations as a reference point, with the goal of making it easier for them to mark up contracts.
Building these kinds of solutions requires keeping up with the machine learning field, which has changed exponentially in just the past 7-8 years. The revolution of generative AI means we see yearly advancements.
For example, when the large language model revolution started five years ago, we transitioned from running code and software on commodity hardware (such as a laptop) to specialized hardware. At the same time, we had to adopt new tools to leverage that specialized hardware, particularly when it came to huge models like GPT3+ that require lots of computing power.
We have to figure out which tools to use to take advantage of these technology advances and best solve the problems at hand.
How would you describe machine learning to someone unfamiliar with it?
There’s a lot of terminology out there: machine learning, artificial intelligence, predictive analytics. But fundamentally, it boils down to the same idea.
I define machine learning as the process of learning patterns from historical data and making predictions or decisions based on those patterns. In the context of Ontra, that could be as simple as predicting which part of a contract is the non-solicit clause based on thousands of non-solicit clauses we’ve seen.
Large language models, like ChatGPT, seem pretty amazing but the premise is simply: “Given the text I’ve seen so far, can I predict the next most likely word?”
Again, it’s learning historical patterns of how words fit together. After the machine has learned from trillions of word sequences, it knows what those patterns look like and can predict what the next word in that sequence should be. It’s a simple idea but there’s a lot of complexity happening behind the scenes to make that work in practice.
How do machine learning and the data-driven products and services you help build add value at Ontra?
At Ontra, our team is involved in different areas with a focus on how to accelerate and streamline the negotiation of routine contracts with machine learning.
At the beginning of a contract, we want to provide the lawyer with the most salient information so they can negotiate as quickly as possible. That might mean automating some data extraction and document classification upon contract intake.
Once a contract is ready for negotiation, we can suggest precedent information or the right reference data to accelerate the mark up. At the end of the negotiation, we can automatically summarize the terms of the agreement and send a report to the customer.
Some of these are ideas and some are features we’ve developed. Whenever we are considering a new feature, we pilot it with the most relevant users, such as those most likely to benefit from the feature.
What’s your favorite application of machine learning in general and at Ontra?
I’ve always been interested in human languages, which is why I got into natural language processing. So, I love machine translation in general. That’s translating between human languages, such as from English to Spanish. It’s a hard task for humans to learn a foreign language but machine learning can translate pretty well. Large language models plus generative AI even makes it possible to transition between languages, such as asking a question in English and getting an answer in German.
At Ontra, I’m most excited about what we can do to accelerate negotiations by following the lead of the latest large language models being trained with instruction following.
When lawyers negotiate a routine contract, they follow a playbook or parameter sheet – essentially instructions for how to run the negotiation.
Once all high-quality data in these playbooks and parameter sheets is digitized, it opens possibilities for how to accelerate negotiations. For example, what are all the ways we can automatically mark up a contract?
How do you stay up-to-date with the latest advancements in the field of machine learning?
I subscribe to several email newsletters for a roundup of the latest news in the field. I also follow the research blogs of interesting tech companies that detail what they’re working on with AI.
In addition, I make time to experiment with the latest open source code and models. The latest advancements are mostly open source, so anyone can download models and run them to see how they work. I learn by doing, so that’s really helpful.
Finally, at Ontra, our machine learning team gets together every other week or so to discuss a research paper. We take turns picking a paper and summarizing it, and then discuss and sometimes run code to experiment with the ideas presented in the paper. It’s a fun way just to exchange ideas and keep up with advancements. Plus, one idea from a paper proved useful in practice, allowing our team to better identify where data is low quality. For example, when a lawyer marks a sentence as a non-solicit clause but it isn’t.
What do you enjoy doing outside of work?
I’ve always been very active, even competing in track & field at university. I was a middle distance runner then, and have dabbled in trail running recently. That said, I rock climb more than I run today. Though I mainly climb in gyms, I’d like to get outside more.
