Dylan Lewis

• Transformed a manual Python-based pipeline into an automated, dockerized Kafka consumer microservice, enabling reliable and scalable generation of a core company product upon API-triggered requests. • Built a cloud-native chatbot application powered by Retrieval-Augmented Generation (RAG) using LLMs, deployed on AWS using Lambda and Bedrock, and orchestrated via Serverless and Terraform. • Provisioned an MSK Kafka cluster and NestJS Kafka Consumer microservice to offload conversation analytics to a dedicated Kafka consumer, enabling parallel processing and preserving chat latency. • Developed an internal full-stack web application to streamline client onboarding and product generation workflows, leveraging NestJS for the backend and HTMX + Alpine.js for responsive, lightweight frontend interactions. • Enhanced observability of several core applications using Datadog SDKs in TypeScript and Python to forward application logging and enable application performance monitoring (APM) • Researched and designed a novel algorithm (patent-pending) to address data robustness and data availability challenges in algorithms dependent on real-time sensor data. Conducted extensive analysis of existing solutions to inform a proprietary design, which was submitted for a USPTO patent for its technical innovation and commercial potential across multiple sectors and applications.

• Designed system and constructed the backend for a web app which assists crisis managers during crisis events by leveraging machine learning models to classify crowdsourced crisis text and image data using TypeORM, PostgreSQL, RDS, NestJS, Docker, Terraform, and AWS namely S3, ECR, Lambda, API Gateway, and SageMaker • Utilized insights from crisis managers in the US and Fukuchiyama (FC), Japan to develop the Human Risk/No Human Risk classification task and determined that the F2 performance metric strongly aligned with their priorities; achieved a 92.8% F2 score on a test set of past flood event Japanese text data in FC with a SVM model and pretrained BERT embeddings • Built pipeline and visualization tool for experimenting with various featurizations of Japanese crisis text data, dimensionality reduction techniques, and clustering algorithms, in order to yield human-interpreted labels from the documents found in each cluster; uncovered 9 human-interpreted labels from past flood event data in FC • Led annotation effort forming a ground-truth test image dataset from past flood events in FC; achieved a 82.5% weighted F1 score on the test dataset using a trained CNN image Flood/Not Flood classification model • Developed an open-source Python package for training, testing, and predicting with pretrained CNNs for classifying crowdsourced crisis image data. See https://pypi.org/project/url-image-module/0.27.0/ • Developed an open-source Python package for featurizing crisis text data, training and testing with a variety of classification machine learning models, and visualizing clusters of featurized text data. See https://pypi.org/project/url-text-module/0.6.1/ • Defined and supervised 8 undergraduate research projects adjacent to thesis research • Presented main findings from research to key stakeholders from a wide range of technical backgrounds

Graduate Teaching Assistant for 6.170: Software Studio • Led recitation sessions and office hours covering the fundamentals of software design and full-stack web development • Mentored project teams providing feedback to students as they developed their final project web applications • Created problem sets for students to practice and solidify concepts taught in lecture and recitation • Graded and provided feedback on students submissions for problem sets and final project milestones • Received an average rating of 6.8/7 by students for stimulating their interest in the subject and showing thorough knowledge of the subject material and 6.9/7 for supporting student learning

• Developed a full-stack web application with React, Redux, TypeScript, Google Protocol Buffers, and CouchDB • Utilized Docker for a containerized development environment as well to build a shareable image of the web application • Designed UI/UX of the application by iterating on the React-Redux frontend based on feedback from peer review

• Translated business logic and user stories into enhancements to a popular car rental website UI using React components • Stylized webpages with SCSS based on design specifications and mockups • Participated in code review to ensure code quality and standards as well as wrote manual tests to ensure that UI updates produced expected behavior

• Led introductory chemistry recitations for 11 incoming MIT freshmen in the Interphase EDGE summer program where students take four college courses, live on campus, and learn about campus resources. • Created worksheets, problem sets, review materials, and exams that assisted students in developing the problem-solving skills they would need to succeed in the chemistry General Institute Requirement. • Acted as a conduit between students in my recitation and the instructor of the course, the coordinators of the Interphase EDGE program, and the teaching assistants for other courses in the program.

Thesis: Towards Automating Crowdsourced Crisis Report Assessment for Enhanced Crisis Awareness and Response