L2 Labs

L2 Labs is an early-stage NYC startup working on data interoperability in the EDI space. As a summer intern my role was twofold: I both supported the research mission of L2 Labs by participating in and designing experiments while also working with the founding engineer on the development of our flagship product.

Research

While working at L2 Labs I worked with the research team to help them solve a schema matching problem between ACORD XMLs and a target vocab. Our problem and the designed solution were formalized and published at CIKM 2026 (Linked here).

While I helped support the entire research process, my main contribution to this paper was the curriculum agent step. In order to bridge the gap between the top-1 and top-10 accuracy of our embedding model, I designed an agentic loop where during training you have a reranker agent that attempts to pick the correct top-1 and then because you have the correct ground truths you have a teacher agent who goes through and makes a curriculum that can be used at inference. This process helped increase our top-1 by 4.7% and beat the SOTA LLM reranker by 12%.

Participating in this process and our research meetings helped me gain valuable experience and knowledge in both machine learning engineering and the research process as a whole. I saw the development of a paper from the first problem statement to the publication and was able to assist with every step in between, giving me valuable experience in how to take an issue from a thought to a published paper.

Development

Along with supporting the research team, I also spent time working on the development of our flagship product Alloy. Along with helping with bug fixes, the main thing I built was a full workflow that took in XML documents and integrated with the tools of one of our customers. This work involved deploying the matcher described in the research above, creating a new UI/UX experience and adding custom business logic for our customer.

While working on the development of this new workflow, I also led a 10+ person user group to get feedback on our design decisions. This session was very beneficial in helping us get a user-friendly product and taught me a lot about how to run A/B testing and user groups in general. Along with this, to support the custom integration, I also met with our customers weekly to get feedback about their use. In the end, after all of the development, we saw a 4x increase in this workflow by our customers, showing that the new workflow gave them better support.

Future Work

After my work this summer I will be continuing to work at L2 Labs part-time while finishing up school, allowing me to continue supporting development and research.