inspiration
i work in a robotics research lab at university of arizona. spent a lot of time watching grad students and professors stare at datasets for weeks trying to figure out what patterns actually matter and what experiments to run next.
the data collection part is solved. we have more data than ever. the hard part is the creative leap – going from "here's what i observed" to "here's what might actually be happening."
when i saw gemini 3's reasoning capabilities i thought – what if we could accelerate that? not replace scientists, but give them a tool that can look at their data and suggest hypotheses they might not have considered.
what it does
you upload a dataset, describe what you're researching, and eureka generates testable hypotheses with full reasoning chains.
each hypothesis comes with:
- the actual hypothesis statement
- step by step reasoning showing how it got there
- specific evidence from your data
- how you'd actually test it
- novelty and confidence scores
the reasoning chain part is important. you can actually follow the logic and decide if it makes sense, not just blindly trust the output.
how i built it
react frontend with vite and tailwind. wanted it to feel like a real research tool, not a hackathon demo that falls apart.
the core is prompt engineering. i spent a lot of time figuring out how to get gemini 3 to think like a scientist – analyze the data structure first, look for patterns and anomalies, then form hypotheses that are actually falsifiable and testable. not just "this correlation exists" but "here's why this might be causal and how you'd prove it."
used papaparse for csv handling and framer motion for the ui animations.
challenges
getting consistent structured output was rough. gemini would write beautiful explanations but the json would be malformed. had to add explicit schemas and fallback parsing.
also had to tune the creativity. too conservative and it just states the obvious. too wild and the hypotheses are unfounded. ended up adding a dial so users can control that tradeoff themselves.
what i learned
prompt engineering for reasoning tasks is completely different from conversational prompts. you need structure and explicit frameworks or the model meanders.
also – gemini 3 is genuinely good at maintaining logical coherence across long chains of reasoning. that surprised me.
whats next
- pdf upload so it can read existing papers and build on them
- citation linking to connect hypotheses to published research
- export to lab notebook tools
Built With
- 3
- ai
- css
- framer
- gemini
- generative
- javascript
- motion
- react
- sdk
- tailwind
- vite
Log in or sign up for Devpost to join the conversation.