United States
Experienced machine learning engineer with a foundation in NLP https://github.com/janest1
Agentic AI, machine learning
Led the development of a semantic cache embedding layer using OpenSearch to enhance LLM chatbot performance, driving a 30% increase in customer acceptance rate and 60% drop in off-topic responses, as well as significant latency and cost savings. Conducted A/B testing to determine caching strategy. Designed and productionized a model training pipeline and predictive model to identify high-quality chatbot responses likely to receive positive customer feedback. These predictions were used to inform LLM prompting strategy and semantic cache coverage, leading to a significant decrease in dropped conversations. Built a scalable text ingestion pipeline with AWS SQS and Lambda for real-time document indexing used in RAG and NER tasks, cutting latency by 50%. Partnered with PMs and infra engineers to shape the product roadmap for AI-generated chat replies and led ML model iteration cycles
-Designed and implemented Argo workflow that periodically fetched chat messages from Snowflake database and applied a clustering algorithm to unclassified messages in order to leverage them as additional training data to improve the accuracy of Drift’s internal topic classification models. Named in patent US-12057106-B2: Authoring content for a conversational bot -Integrated AI model predictions and conversational data into Drift’s customer engagement scoring feature by building an API service and dynamoDB table to serve as a real-time feature store. This introduced AI to a new part of the product while adding minimal complexity to the engagement score calculation.
Investments AI -Built tools to facilitate large-scale text annotation efforts for AI-enhanced document viewer -Designed and implemented end-to-end pipeline that retrieved annotated documents from external crowdsourcing providers, calculated quality metrics, and transformed text into training dataset for deep learning event detection model -Created service that sampled raw text for annotation at document- or sentence-level. The Elasticsearch-based sampler ranked results based on custom keywords in order to balance distribution of labels in training data and improve classification accuracy on sparse data -Conducted quality assurance workshops with external annotation providers and maintained technical documentation to improve quality of labeled text
-Designed and developed linguistic rule-based information extraction system as part of a tool which evaluated the quality of scientific research papers -Extracted and normalized text from academic papers with Python and Java for use in company-wide testing suite
Annotate texts and audio transcriptions to build a corpus of verb-phrase ellipsis