Emily G.

LATE POST : Week 2 Update from the Headstarter AI's Fellowship! As part of the fellowship, I have had to learn an entirely new tech stack in order to be able to proceed through the assigned projects. I’ve been building My Pantry, a web app designed to help you manage your pantry. It’s powered by Next.js, Firebase, Material UI, and the Meta Llama 3.1 8B Instruct API. Key Features: Inventory Management: Keep track of everything in your pantry with ease. Recipe Suggestions: Discover personalized recipe ideas using the items in the pantry with help from Meta AI. Coming Soon: Adding computer vision to automatically recognize and log items using photos. Check it out: https://lnkd.in/dVtYDpa8 While the UI and the features are very basic, I will be enhancing it as I go upwards with the learning. A huge thanks to the Headstarter team for this fantastic learning experience. #Headstarter #WebDevelopment #NextJS #Firebase #MaterialUI #LearningJourney #CodingCommunity

6

2 Comments

✨ Reflections on My SWE Internship in a Non-Technical Team ✨ This summer, I had the privilege of working as a Software Engineer Intern on the Annotations Team at Cohere. This internship offered me a unique blend of challenges and opportunities that helped me grow and learn in unexpected ways. Key Takeaways 🔹 Mastering Prioritization: Balancing smaller feature tasks alongside a major project taught me how to manage priorities effectively, stay organized, and maintain momentum in a fast-paced environment. 🔹 Creating Automation Impact: One of the most fulfilling aspects of my internship was building automation tools that significantly enhanced workflow efficiency. Seeing the tangible impact of my work on everyday processes was incredibly rewarding. 🔹 Bridging Technical & Non-Technical Teams: Communicating complex technical concepts to non-technical teammates was a valuable learning experience. It improved my ability to break down intricate ideas clearly and concisely, which is crucial for cross-functional success. Challenges That Shaped My Growth 🔹 Independent Problem Solving: While working independently meant fewer opportunities for real-time collaboration with engineers, it pushed me to develop my problem-solving abilities and boosted my confidence in handling challenges on my own. A heartfelt thank you to Trisha Starostina and Claire Cheng for being exceptional leaders. Your guidance made my internship not only productive but deeply rewarding. Special thanks to my mentors Charlie Wang and Marina Machado—your mentorship was invaluable in shaping my growth and ensuring I navigated my projects successfully. And a big thank you to Maria Bowlware for a seamless recruitment experience. ✨ What’s Next ✨ I’m thrilled to be rejoining Cohere this fall, this time on the Platform Team, where I'll be focusing on customer-facing model-serving applications. I'm excited to take on new challenges and contribute to the next phase of Cohere's growth! #SoftwareEngineering #InternshipExperience #Mentorship #CareerGrowth #GrowthMindset #Cohere

100

6 Comments

Last week, I wrapped up my internship at 𝗘𝘀𝗿𝗶 𝗮𝘀 𝗮 𝗦𝗼𝗳𝘁𝘄𝗮𝗿𝗲 𝗗𝗲𝘃𝗲𝗹𝗼𝗽𝗺𝗲𝗻𝘁 𝗘𝗻𝗴𝗶𝗻𝗲𝗲𝗿 𝗜𝗻𝘁𝗲𝗿𝗻 𝗶𝗻 𝘁𝗵𝗲 𝗚𝗲𝗼𝗱𝗮𝘁𝗮𝗯𝗮𝘀𝗲 𝗧𝗲𝗮𝗺. I still remember my first day when my mentor, Christopher Curtis, mentioned that I had been assigned one of the most complex projects for the summer. He advised, "Never hesitate to ask questions and interrupt immediately when in doubt." I embraced this advice, conducting research on every topic I encountered and seeking clarity whenever necessary. There was no predefined roadmap to achieve the solution. I took it as a challenge, stayed curious, and started digging deep into the problem statement, trying to understand the architecture of the 𝗔𝗿𝗰𝗚𝗜𝗦 𝗘𝗻𝘁𝗲𝗿𝗽𝗿𝗶𝘀𝗲 𝗦𝗲𝗿𝘃𝗲𝗿 𝗮𝗻𝗱 𝗨𝘁𝗶𝗹𝗶𝘁𝘆 𝗡𝗲𝘁𝘄𝗼𝗿𝗸. The team was delighted when I presented my approach to solving the problem, and once approved, I began the implementation. The foundation and design of the tool, using 𝗢𝗢𝗣 𝗱𝗲𝘀𝗶𝗴𝗻 𝗽𝗮𝘁𝘁𝗲𝗿𝗻𝘀 𝗮𝗻𝗱 𝗽𝗿𝗶𝗻𝗰𝗶𝗽𝗹𝗲𝘀, were crucial since the tool needed to adhere to code quality, feature extension, and scalability. The tool is aimed at drawing conclusions on the correctness and performance of the server. Alongside working on the project, I participated in a two-day-long 𝗶𝗻𝘁𝗲𝗿𝗻 𝗵𝗮𝗰𝗸𝗮𝘁𝗵𝗼𝗻 and 𝘀𝗲𝗰𝘂𝗿𝗲𝗱 𝘀𝗲𝗰𝗼𝗻𝗱 𝗽𝗹𝗮𝗰𝗲. I showcased my leadership skills while also fulfilling full-stack development duties. One of the highlights of the internship was attending the Esri User Conference 2024, where I interacted with industries utilizing Esri services and technologies. I engaged in tech sessions and workshops, including 𝗚𝗲𝗼𝗔𝗻𝗮𝗹𝘆𝘁𝗶𝗰𝘀 𝗘𝗻𝗴𝗶𝗻𝗲 𝗮𝗻𝗱 𝗨𝘁𝗶𝗹𝗶𝘁𝘆 𝗡𝗲𝘁𝘄𝗼𝗿𝗸 𝗼𝗻 𝗙𝗶𝗲𝗹𝗱 𝗠𝗮𝗽𝘀. I am immensely grateful to my mentor and the entire Geodatabase team for their support and guidance throughout the internship. I truly feel fortunate to have been part of such a remarkable team. This experience will forever hold a special place in my heart, and I am eager to carry forward all the skills I learned at Esri into my professional journey. #Esri #SoftwareEngineering #Internship #Geodatabase #ArcGIS #OOPDesign #Hackathon #EsriUC2024 #Leadership #ProfessionalGrowth

57

Last summer, I had the incredible opportunity to explore the exciting field of Extractive Question Answering (EQA) as part of my research at Denison University with Son Tran. Context of the problem: EQA is pivotal in Machine Reading Comprehension, a rapidly advancing field in NLP. Despite its progress, a significant challenge is handling unanswerable questions—queries that can't be resolved with the provided context. These questions are crucial for training robust models, ensuring they avoid misleading answers and perform effectively in practical applications. The Problem: However, manually creating unanswerable questions takes time and effort. Traditional methods like SQuAD2.0 rely on labor-intensive manual annotation, which is expensive. Moreover, models trained on SQuAD2.0 often need help with out-of-domain samples, limiting their real-world effectiveness. The Solution: We developed a novel pipeline that automates the creation of unanswerable questions. Here's how it works: - Matchmaker: We pair existing questions with new, unrelated contexts to create potential unanswerable questions. It's like setting up a blind date. We use TF-IDF to ensure the questions are closely related to the subject matter. - Detective Work: Using a technique called 'adversarial filtering,' we identify the genuinely tricky, unanswerable questions to minimize errors and ensure their unanswerability. - Quality Check: Our filtering model ensures no answerable questions sneak through, maintaining high quality and accuracy. Then, humans validate the final questions, resulting in a low error rate and better than human-annotated datasets like SQuAD2.0. What I've Learned: I gained experience training and inferencing transformer-based models like BERT, RoBERTa, and SpanBERT, which are crucial for natural language understanding and EQA tasks. By leveraging TF-IDF, I've crafted unanswerable questions that accurately match with relevant paragraphs. My experience includes meticulous dataset quality-checking and interpreting model results to maintain experimental integrity. A valuable lesson on RoBERTa's sensitivity to whitespace—after inadvertently destroying a 16-hour experiment—taught me the importance of precise data preprocessing. Moreover, I've emphasized writing clean, well-documented code to enhance clarity and maintainability, which is crucial for optimizing our pipeline's efficiency and contributing to open-source projects. Implications and Results: Our research introduces a pivotal advancement in machine reading comprehension within our pipeline. It enhances EQA model performance and addresses critical challenges in real-world applications, laying the groundwork for more reliable and versatile machine comprehension systems. Special thanks to Son Tran for his invaluable mentorship and support throughout this journey. He made this summer the very best! For insights into our novel pipeline, check out our research paper: https://lnkd.in/eWhxkxCp.

31