SurvAI

Inspiration

Hackathon feedback forms often feel long and repetitive. People either rush through them or give vague answers which makes it hard for organizers to identify what actually worked and what needs improvement. We wanted to build something that feels more like a game than a survey while still producing clearer and more constructive results.

What it does

ServAI is an enjoyable AI adaptive survey that asks the next best questions in real time from the number of questions given by the survey creator based on their previous responses. As the user answers, the system estimates overall sentiment across five experience categories specified by the survey creator (e.g Very Satisfied, Very Unsatisfied) and updates a certainty score. Once the confidence is high enough (or a maximum question limit is reached), ServAI concludes the survey and displays its final guess of which category the user most likely aligns with.

How we built it

We built a React + Vite frontend. On the backend we used FastAPI in Python to manage sessions, store question banks, and compute a running probability distribution across five sentiment categories. We integrated the OpenAI API to infer sentiment for each free-text answer, then combined those results using a weighted posterior model that accounts for confidence, recency, and contradictions. We also added hardcoded keyword-based boosting to make certainty respond more strongly to explicit positive/negative language.

Challenges we ran into

The hardest part was balancing “fun and fast” with “accurate and trustworthy.” Certainty tuning was the most difficult part. we needed the model to converge quickly without becoming unstable or overconfident. We also ran into integration issues where the frontend accidentally hit the wrong server and received HTML instead of JSON, which caused the UI to display broken “questions.” then we had to optimize latency and caching so each answer didn’t trigger unnecessary repeat API calls.

Accomplishments that we're proud of

We delivered a complete end-to-end system, importable question banks, session based adaptive routing, AI-based sentiment inference, live confidence tracking, and a clean finish screen that summarizes results transparently. We’re proud that ServAI turns a traditionally boring process into an engaging interaction while still producing structured output that organizers can use immediately.

What we learned

We learned how to design an adaptive questioning loop, how to combine probabilistic scoring with real-time UX, and how sensitive “confidence” systems are to parameter choices. We also learned practical lessons about API reliability (timeouts, retries, caching), frontend backend contract design, and building user-facing transparency so the final output feels credible rather than random.

What's next for ServAI

Next, we want to improve question selection beyond sequential order by choosing questions that maximize information gain based on current uncertainty. We also plan to add analytics dashboards for organizers, exportable summaries, and stronger guardrails against adversarial or low-effort responses. Longer term we want ServAI to generalize beyond hackathons so any event or organization can run a gamified adaptive survey that reaches confident insights with fewer, smarter questions.

Built With

• fastapi
• javascript
• python
• react