Built YC-backed ML-based startup that integrated sales call recording solutions and CRM software with state-of-the-art NLP solutions to autonomously monitor call recordings for salient product and market trends. Served 14 enterprise customers, processing over 50K conversations and enabling 1K+ customer-facing insight reports, and raised VC funding from Y-Combinator, Soma Capital, Maschmeyer Group Ventures and a litany of industry operators and angels.Leveraged by marketers & product managers, Buzzle provided users with autonomous trend detection and reporting, helping leading companies revise marketing strategy, define product roadmaps and discover revenue blocking customer pain points.Formed partnership with the Montreal Institute for Learning Algorithms led by Dr. Yoshua Bengio (Turing Award Winner).Built and managed entirety of application backend through resilient, cloud deployed components such as: - GraphQL API: Leveraged AWS Appsync and Apollo client to deploy custom Lambda backed resolvers.- Vector Databases: Leveraged Pinecone to store high dimensional vector data for complex natural language tasks.- REST APIs: Deployed and scaled REST APIs (FastAPI, ECS) fronting a pool of worker instances (ECS) in a workqueue system (Redis) to service machine learning tasks (Embedding, Classification, Topic Understanding).- Other technologies: Serverless microservices, Elasticsearch clusters, SQL & NOSQL databases, etc.Trained ML models and pipelines providing industry-leading trend detection and topic understanding. Deployed models in a self managed compute cluster capably delivering on machine learning tasks such as:- Sentence Embeddings via a custom fine tuned language model.- Classifiers for supervised ML tasks such as complexity detection, usefulness detection and sentiment detection.- Novel techniques for knowledge graph exploration and topic understanding.

Led team of ML engineers, software engineers and interns building internal NLP solutions.Deployed NLP solutions across the company delivering value to stakeholders such as Legal, Model Risk, Risk, Cyber, etc:- Delivered real time risk topic monitoring solutions in direct response to US Treasury mandate whilesimultaneously managing internal audits and heightened oversight from the Federal Reserve.- Optimized the model risk governance process by over 300% through automated model white paper scoring.- Improved patent acceptance and placement rates via internal patent review tooling.- Expedited critical incident detection times (by greater than 5%) through real time sentiment anomaly detection of company social channels.- Rebuilt internal call center reason classifiers for a 23%-increase in accuracy and multi-class capability.Co-lead “Tiger” team which designed, built, and deployed internal machine learning as a service platform.- Platform extended the open source implementation of Prefect to provide DAG based orchestration and delayed compute capabilities via Dask and Dask Gateway.- Implemented dynamic runtime packaging to ultimately enable the composition of custom models in a DAG and promote reuse of common machine learning workflow patterns at Capital One.- Created a consortium of connectors to enterprise data stores (streaming, Snowflake, S3, etc) and designed a REST API for platform use with a client facing SDK for optimal modeler experience.- Deployed platform in a Kubernetes cluster with multi region resilience and automatic failover.Mentored and taught intern teams to complete production NLP deliverables and future placement at FAANG & YCombinator.

- Working on Machine Learning as a Service (MLaaS) as a member of the Capital One Center for Machine Learning (C4ML) facilitating model orchestration at scale - Building Natural Language Processing models delivered through the MLaaS platform.Key Responsibilities:- Develop NLP models supporting Model Governance, Regulatory Risk, Retail/Commercial Call Centers, and more.- Lead & mentor UIUC/UMD intern teams.- Collaborate on innovative model orchestration system facilitating streamlined, DAG based model pipelines.

- Leveraged machine learning techniques to build software tools for the cloud engineering organization- Built a slack based chatbot to assist support teams through automated, ML powered question answering leading to a reduction in engineer time spend and a dramatic improvement in response times. - Constructed "Hygiene" tool - utilized to identify defective cloud builds and eliminated the turnover of poorly constructed cloud resources. - Lead group of interns in designing "CloudCricket" - an instant alerting application designed to detect and provide immediate alerting on problematic changes to Capital One's cloud environment.

- Published original research at NIPS 2018 as part of the AI for Social Good workshop (https://aiforsocialgood.github.io/2018/pdfs/track1/108_aisg_neurips2018.pdf)- Focused on two core paradigms: 1) Let students tutor our agent with their voice. 2) Focus on knowledge parallelization on areas of weakness.- Currently building application off of research for deployment within classroom pilot programs (https://sites.google.com/view/qai/home)- Partnered with Mila (Montreal Institute of Learning Algorithms) as part of their Social Good AI initiatives. - Building and deploying state of the art models for upstream machine learning tasks.

- Assisted Dr. David Chesney in teaching “Software Engineering” – a course with a focus in development for social good- Helped students with the development of projects with focuses in augmented reality telemedicine, prosthetic cyber manufacturing, and accessible natural language processing.

- Led robotics team focused on development of autonomous mining rover for the NASA Robotic Mining Competition.- Developed Kinect based obstacle avoidance utilizing point cloud subtraction. Smoothed data through Moving Least Squares Approximation and clustered obstacles utilizing Euclidean Cluster Extraction. - Incorporated LiDAR-centric localization by detecting intensity values from an AR pattern and segmenting edges to accurately understand rover pose. Supported redundantly with RGB camera detection of April Tags.https://rest.engin.umich.edu/

- Developed SVM model to detect anomalies in internal log files leading to a dramatic reduction in fraudulent false positives.- Utilized active learning built upon the HintSVM model to efficiently use large unlabeled dataset – extracting historic fraudulent examples from existing Salesforce logs and labelling non-anomalous data appropriately. - Constructed production ready model and network architecture on the cloud through AWS, InfluxDB, and Grafana.- Built Alexa skill facilitating access to the Capital One rewards program. (Winning Intern Hackathon Project)