RE:TRY

Inspiration

We were inspired by a simple problem: we get a lot of advice, but very little help figuring out whether that advice actually works for us. When someone struggles with procrastination, focus, studying, routines, or other everyday challenges, the usual response is to immediately recommend a solution. But the same solution doesn't work for everyone, and we often don't understand the real reason behind a problem in the first place. We wanted to explore a different approach: what if AI didn't immediately tell you what to do, but helped you investigate the problem first? That idea became RE:TRY; a tool built around curiosity, experimentation, and learning from real-world results.

What it does

RE:TRY turns an everyday struggle into a small investigation. A user starts by describing something they're struggling with. Instead of immediately receiving advice, RE:TRY asks focused questions to uncover patterns, triggers, context, previous attempts, and exceptions. Once the investigation is complete, RE:TRY generates three plausible hypotheses about what might be contributing to the problem. Each hypothesis includes the evidence behind it, what remains uncertain, a confidence level, and a suggested experiment. The user chooses one hypothesis and tests it through a small real-world experiment. Afterwards, they record what happened and reflect on the result. The core loop is: Investigate → Hypothesize → Test → Measure → Learn → RE:TRY The goal isn't for AI to decide what's right for someone. It's to help them make better observations about their own experience.

How we built it

We built RE:TRY as a web application using React and Vite for the frontend and Node.js with Express for the backend. The frontend manages the investigation flow, conversation history, hypothesis selection, experiments, and results. For the AI layer, we connected the backend to the OpenRouter API. We designed separate AI prompts for two major stages: Investigation: generates one focused question at a time based on the user's problem and previous answers. Hypothesis generation: analyzes the investigation and produces three evidence-based possibilities while explicitly communicating uncertainty. We also added safeguards to keep the system from presenting hypotheses as diagnoses or facts. Git and GitHub were used throughout development to track and manage the project.

Challenges we ran into

One of our biggest challenges was getting the AI to behave like an investigator instead of an advice generator. A normal chatbot naturally wants to answer a user's problem. RE:TRY needed the opposite behavior: it had to resist jumping to conclusions, ask only one question at a time, use previous answers, and uncover information that was still missing. We also had to deal with AI reliability. At one point, an authentication/environment-variable issue caused our OpenRouter requests to fail. Because the application had fallback questions, the failure initially appeared as the AI repeatedly asking the same question. Debugging the chain from frontend → backend → API helped us understand how important graceful error handling and observability are. Finally, we had to make difficult decisions about scope. There were many features we could have added, but we learned that getting the core investigation-to-experiment loop working was more valuable than building a huge feature list.

Accomplishments that we're proud of

We're proud that RE:TRY is more than a chatbot with a different interface. We built an actual reasoning workflow around the AI. Instead of: User → AI answer RE:TRY creates: User → Investigation → Evidence → Hypotheses → Experiment → Result → Learning We're particularly proud of the hypothesis system because it deliberately separates evidence, uncertainty, confidence, and testing rather than presenting an AI-generated explanation as truth. We're also proud that we turned an abstract idea into a working end-to-end product within the hackathon — from the frontend experience to the backend AI integration and experimental workflow.

What we learned

We learned that building with AI isn't just about getting a model to produce impressive responses. The product design around the model matters just as much. A powerful model can still produce an unhelpful experience if the prompts, workflow, and constraints aren't designed carefully. We also learned to treat AI outputs as hypotheses rather than unquestionable answers. Designing RE:TRY forced us to think more carefully about uncertainty, evidence, failure cases, and responsible AI. Most importantly, we learned that sometimes the best thing an AI can do is ask a better question instead of giving an immediate answer.

What's next for RE:TRY

The current version is only the beginning. Our next step is to make RE:TRY capable of learning from multiple experiments over time rather than treating each experiment as an isolated session. We want to build a personal experiment history where users can see: What they've tried What happened Which hypotheses received support Which were ruled out What patterns keep appearing What they might test next We also want to improve the quality of experiments by making them more measurable and personalized, while strengthening safety boundaries around sensitive wellness topics. Eventually, RE:TRY could become a broader personal experimentation platform; helping people approach everyday problems with curiosity and evidence rather than immediately assuming they know the answer. Because sometimes the most useful answer isn't an answer at all. It's a better experiment.

Built With

• express.js
• node.js
• open-router
• react
• vite