training big neural networks @ cohere

My team focuses on applying state-of-the-art research from the Natural Language Processing (NLP) research community to various conversational commerce products on a global scale. I work with partner teams to develop AI products from initial definition through GA release. I’m responsible for my team’s delivery of machine learning research that supports flagship AI products across the LivePerson Conversational Cloud.To this end, we leverage state-of-the-art transformer architectures such as BERT and related variants to train language models from scratch using one of the largest conversational data sets. We support numerous downstream models that underpin Tier 1 services. To deliver low latency inference speeds in high SLA environments using these large language models my team implements state of the art model optimizations such as distillation and quantization. We further utilize these BERT high-dimensional representations to augment end-user domain knowledge via unsupervised and semi-supervised modeling (TSNE, UMAP, KNN, etc.) techniques.In addition to the applied machine learning described above, we work on more basic research in the following areas:- Natural Language Generation (NLG)- Probabilistic Dialogue Management (DM)- Dialogue Quality Measurement- Conversational Language Model Pretraining using self-supervision

Lead the machine learning group focusing on products developed using state-of-the-art natural language processing techniques.

We are hiring for a range of science and engineering roles. If you have experience and interest in NLP/NLU, Machine Learning, ML engineering, and ML service building, please PM me! Implemented unsupervised models in Pytorch and Tensorflow. Built ETL and training pipelines. Built visualizations to support data exploration and iterative model training for goal oriented conversational models.

As part of CILVR at NYU, I worked with Prof. Kyunghyun Cho to develop new data sets and models to solve open domain Question and Answer tasks. Ported a SOTA architecture from Theano to Pytorch.

Designed, built, and maintained:- Visualization tool for comparison of >1,000 bladder cancer patients and >500 somatic DNA mutations.- 3rd order polynomial data correction model for DNA methylation status in bladder cancer patients. - Patient records and reports, including demographics, medical history, and biomarker profile.- Order and human sample management solution.- Custom integrations for onsite and cloud systems (78% below budget and 5 months ahead of schedule).Managed technical sub-contractors.

Designed and built systems longitudinal client tracking.Built visualization tools for analysis of low signal-to-noise data.Managed a team of four consultants across multiple projects.Client facing project lead on multi-million dollar SFDC implementations.

Built platform to facilitate energy retrofitting marketplace. Built platform to facilitate bulk energy auction management.