NodeBud

Problem

Practicing for behavioral interviews is awkward. You can rehearse in front of a mirror, but you never get useful feedback on the parts that actually matter — whether you held eye contact, whether your posture read as confident, whether your answer rambled or stayed on the question. Built in 36 hours at BroncoHacks 2026.

Approach

NodeBud (CareerPrep AI) runs a full mock interview in the browser. The user picks a role, answers prompts on camera, and gets per-question scoring plus aggregate feedback at the end. Three feedback channels run in parallel:

• Posture and eye contact — webcam analysis tracking whether the user is facing the camera and holding gaze
• Voice answer scoring — speech-to-text on the response, then evaluation against the question
• Session history — persistent progress across multiple practice runs

All processing happens locally — video never leaves the browser, which kept the privacy story simple for the demo.

Tech

• Next.js + TypeScript — frontend and routing
• React — UI components and state
• Browser ML APIs — pose / face detection in-page
• Vercel — hosting

Outcome

Shipped a working demo at the deadline with the three feedback channels wired end-to-end. The hard parts were timing — getting the webcam pipeline, the audio transcription, and the scoring loop to all stay in sync without freezing the page during a question. Most of the second night was profiling and trimming the model loads so the app stayed responsive on a typical laptop.