Gradly

Inspiration

Research papers contain valuable ideas, but reading them can feel overwhelming. Summaries help people get through the text, yet they offer little opportunity to practice interpreting evidence or explaining a finding.

We built Gradly around a question: what if a research paper became something you could actively learn through? Our goal is to help learners move from reading a conclusion to understanding how the evidence supports it—and where that evidence stops.

What it does

Gradly turns research papers into evidence-linked learning challenges.

Learners can explore prepared studies or upload a PDF and review its extracted text. The learning experience combines study briefings, contextual research terms, interactive questions, and opportunities to explain their reasoning. Evidence review brings learners back to the source, while transfer scenarios ask them to apply an idea in a new situation.

Saved progress lets learners continue their work, and practice points recognize completed challenges. These points track practice rather than certify mastery.

Prepared examples and sample previews are clearly distinguished from analysis of an uploaded paper.

How we built it

Gradly uses React and Vite for the learner interface, TypeScript across the application, and an Express backend. Shared contracts connect the frontend and API, while reusable lesson-state logic manages attempts, progression, and rewards.

The research service supports PDF text extraction, OCR, processing review, and saved learning history.

We integrated NVIDIA Nemotron through the Nebius Token Factory API for reasoning assessment and research lesson generation. The reasoning-feedback route is configured for Nemotron 3 Super, with configurable models for generation stages.

Model calls and credentials remain on the server. Trusted evidence, bounded prompts, and response validation help keep feedback grounded. When inference fails, the application shows an explicit failure state instead of silently substituting a sample response.

Challenges we ran into

The hardest challenge was keeping the learning experience faithful to the research. A fluent AI response can still overstate a finding, overlook a limitation, or ask a question the paper cannot answer.

We worked to distinguish source evidence, authored teaching content, model output, and fictional practice scenarios. We also adapted generation prompts for different research approaches, rather than assuming every paper should be taught through statistical tests.

PDF extraction and longer processing steps created additional challenges. Extraction review, resumable progress, retry handling, and duplicate-reward protection helped make the experience more reliable.

Accomplishments that we're proud of

We brought paper intake, guided practice, contextual help, and evidence review into one learning experience.

We are particularly proud of the attention to evidence boundaries: keeping assessment criteria out of the initial learner payload, labeling sample content, preserving answers after failures, and preventing repeated submissions from duplicating practice points.

Gradly demonstrates a practical use of AI beyond summarization: supporting a learning activity in which people make decisions, explain their thinking, and inspect the evidence.

What we learned

We learned that learning design requires as much attention as model integration. A useful challenge needs sufficient context, a clear purpose, and a defensible connection to the paper. Feedback needs to help learners examine their reasoning.

We also learned to separate a functioning prototype from proven educational impact. Software checks can verify progression, retries, and response handling. Establishing whether Gradly improves understanding requires evaluation with learners and educators.