Featured Student Projects

During my research visit to the Bay Area, I met with over a twelve researchers and practitioners across Stanford and UC Berkeley working in human-computer interaction, AI ethics, digital mental health, trust and safety, education, and online communities. My conversations included Stanford faculty and researchers such as Nicole Martinez-Martin, Jared Moore, Myra Cheng, and Merve Cerit, as well as Berkeley scholars Abigail De Kosnik, Jasmine Yen-Ching Shih, and Samia Menon. I also attended Stanford's Human-Computer Interaction Seminar and NLP Seminar and spoke with Dave Willner, Bowdoin alumn and industry professional working directly on AI safety and platform governance.

This experience helped me gain a wide range of perspectives on emerging AI systems. Researchers discussed topics including youth interactions with AI companions, digital well-being interventions, trust and safety infrastructure, content moderation, AI evaluation, consent and privacy, mental health applications, and methods for studying human-AI relationships using both computational and qualitative approaches. Several conversations focused on practical research challenges, including working with chat logs, designing studies with young adults, navigating IRB processes, collaborating with industry partners, and translating academic findings into policy or product recommendations. Meeting with faculty, postdoctoral researchers, and PhD students at different stages of their careers also provided valuable insight into academic research pathways and opportunities for future collaboration!

Screenwriting is difficult, it’s hard to learn, hard to use software, and expensive. That’s why I decided to make an easy-to-use beginner friendly screenwriting software! Inside Draftroom, a user can sign up, create a project, and write a screenplay in industry-standard format. If you don’t know how to write scripts, have no fear, head over to the help center. Inside the help center, there are tutorials for mastering format, structure, and craft. For PRO users, it even comes with practice and quizzes at the end of each module. The site auto-formats any piece of writing—scene headings, parentheticals, dialogue, characters—and you can use keyboard shortcuts for ease. Once a script is built, the user can generate a beat sheet (if you don’t know what a beat sheet is, head over to the help center inside the website to learn, under Module 3!) just from the script itself. Users can also create a storyboard, generate a list of the Cast and Crew with stats for each, and keep track of costs for the budget. My personal favorite feature of this app is the Schedule. For those who have a 100-page script and 50+ scenes, you understand how difficult it is to break everything down into what to film and when. DraftRoom automatically separates all of your scenes, groups them by location and time, and creates a shooting schedule. No more freaking out after you realized you needed one more shot at the park that’s half an hour away.  If you have the Pro subscription, you can do all of this with the help of AI. It’s like having a writing assistant always with you. The AI can do basically anything you want it to within the script. After you write your own prompt or click suggested prompts—Find the dead weight, Find on-the-nose dialogue, Map the protagonist's objective, Scene-by-scene outline, Write 5 taglines, Opening title card—it automatically edits your script.

Claude Code is not inherently easy. It initially seems quite simple to use, as it seemed to be coding without “coding,” but there were many different failures and learning opportunities that presented themselves throughout this entire project. One of the most difficult things for Claude to understand was the script's format/page dimensions. It seemed simple to ask it “make sure the page dimension is 8.5x11,” but that could not be further from the truth. It would not listen. The only way that I made it work was telling Gemini about my issue and making it create a prompt to fix the pages. Another difficult task was making sure that my AI API key was stored in a safe environment while also having no idea what to do. As soon as I would post my API key into Claude Code, it would freak out and tell me to immediately change it because you shouldn’t share it with anyone. This turned into a valuable lesson about web safety with Claude, it’s never guaranteed. Even if you ask Claude to make everything safe, human error can compromise a site, especially if you don’t understand where to put keys or how to turn environmental variables to “secret.” The most difficult task was making my website look good, and I'm not sure I succeeded. I tried my best implementing third-party designs from 21st.dev, but I still found it almost impossible to make the site not look vibecoded. I ended up telling Claude Cowork that it was a "Professional UI/UX Design Analyzer (if that’s even a thing)” and told it to create a spec sheet of everything wrong with it and I fed it to Claude Code after. To be completely honest, it was underwhelming, but it made it look slightly more visually appealing. One thing that most people don’t realize about Claude Code is the “one more prompt” issue. After you have a solid framework, a user will have a list of things to fix/add. At some point this user will get tired and tell themselves, “I’ll only do one more prompt.” This is a grave error. It’s never just one more prompt. An hour since you said, “one more prompt,” the user is in a worse place than they started, adding useless features. This project has been a battle of knowing when exactly to stop due to Claude Code being so accessible. It’s easy just to tell Claude to add another crazy feature you want in your app, knowing when to be content with your project is perhaps the most difficult feature of Claude Code. 

I recently completed an honors project in the Economics Department related to Artificial Intelligence (AI) and sustainability, supervised by Professor Martin Abel. I am very grateful to have been able to use Hastings Initiative grant funding to recruit study participants for my experiment exploring how people update their beliefs and behavior in response to information about the environmental impacts of AI. The mini grant supplemented additional funding I received from various other sources, which enabled me to increase my study sample to reach my target of around 1,400 people, enhancing the precision of my study.

Working on my honors project was an extremely rewarding experience as I gained practical skills in how to design and execute a novel research project. I have thoroughly enjoyed being able to dive deep into a topic of increasing salience and meaningfully contribute to an emerging economics literature in a professional manner. I am deeply appreciative that the Hastings Initiative Student AI Grant enabled me to appropriately address my research question and continue to develop research skills that will be extremely relevant beyond Bowdoin.

My Hastings Initiative Student AI Grant funded a Claude Pro subscription that supported two projects this semester, both of which converged on questions about the relationship between form and argument. The first was APDA Practice Room, a web-based debate simulator I built as my final project for DCS 2470: AI in the World, which generates structured cases, runs full six-speech rounds, and produces judging decisions grounded in my team's institutional knowledge base; it is deployed at apda-practice-room.vercel.app and was developed using Claude Cowork and Claude Code. The second was an independent study in early modern English literature, for which Claude served as an analytical instrument across thousands of pages of seventeenth-century anthologies in search of poems related to an anonymous 1639 piece called "Upon the Letter O" on which my final paper now rests. Without that capacity to read at scale, identifying the period material that gave my paper its critical context would have been prohibitively difficult.

What linked the two endeavors was less obvious at the start than it became in retrospect. Both required Claude to function as a sustained analytical partner against a structured body of material, my team's debate knowledge base in one case and the early modern anthology archive in the other. Both surfaced the same lesson about the relationship between form and argument: in competitive debate as in post-regicide royalist poetry, the most consequential work occurs through arrangement, juxtaposition, and figural inversion rather than through explicit content. The paper itself bore out the same principle, becoming fruitful only after substantial grunt work and organic human thinking; at this level of analysis, Claude could orient the inquiry and provide a starting point, but the analytical labor itself has to be done by humans to have true value. That recognition shaped how my debate team thought about the simulator's deployment. When I demoed the simulator to three other varsity debaters, their reaction was that the tool should not be deployed to novices, since a first-year debater ordinarily develops analytical reflexes through the slow grunt work of producing tight cases, failing, and producing more, and APDA Practice Room collapses that formative labor into a prompt. Our team's resolution, with which I agree, is to restrict the system to its practice and judging functions, where the simulator augments human work rather than replacing it.

I appreciate the Hastings Committee allowing me to use Tripo AI for my final project in AI in the World. I have always dreamed of creating virtual worlds and gothic villages, so I decided to make a virtual 3D gothic village where users could walk around and view the village. It required 3D models to be made, as well as a software that would allow me to place the structures as I pleased. I used the funding to allow Tripo to create 3D models from 2D pictures I had hand selected from Pinterest. The paid version allowed unlimited models to be created for my village for one month, as well as allowed me to export the newly-created 3D models to other platforms. For my project, I uploaded the finished structures to Spatial.io, which allowed me to rearrange the structures into a village and town circle formation, with a castle in the distance. The overall project went smoothly, and I credit Anthropic’s Claude for helping me navigate new software that I had never used before to get the results I wanted.

I would recommend Tripo to future students. It was very intuitive and easy to navigate, and it also didn’t require any coding. Spatial.io also didn’t require any coding. For students who would like to make a more personalized and detailed world, I would recommend maybe City Engine. I would say that although Spatial was limited in what I could do, it worked well for this project because I just wanted to insert my buildings into a digital world.