Overview

Scouter is a proctored assessment platform for finance recruiting. Recruiters send candidates a timed case study, such as a financial modeling test, which candidates complete in the browser while their webcam and screen are recorded. An AI pipeline then flags suspicious moments, like reference material on screen, messaging apps, or a second person in frame, so recruiters review flags next to synchronized video clips instead of watching hours of footage.

Stack: React + TypeScript, Node/Express, PostgreSQL (Supabase), Google Gemini through LangChain for frame analysis and grading, BullMQ + Redis for background video processing, and browser-native MediaRecorder for simultaneous webcam and screen recording.

Key Contributions

AI grading dashboard: Built a new Grading tab (~4,700 lines) where recruiters upload candidate submissions and see Scouter's AI rubric grade, including the reasons each section lost points. Teams add their own scores and comments per section and record a High pass / Pass / No pass decision. Rolled out behind an email allowlist, with follow-up work on concurrency edge cases such as stopping stale writes and holding grading slots until grading finishes.
Face-framing check: Added in-browser face detection with MediaPipe so a candidate's face must be fully in frame, with exactly one person visible, before they can accept the proctoring terms. Detection runs on the same cropped view the candidate sees, and was hardened with a timeout for hung detectors, retention of slow-but-successful loads, and camera permission required first.
File-deletion verification: After submission, candidates must delete the case-study files they downloaded, and AI checks that they did. Fixed cases where the check could falsely reassure recruiters, such as a session with mostly failed frame loads still scoring as fully checked; a coverage threshold now marks those sessions "unverifiable". The deletion list now shows real download counts and warns about duplicate copies.
Document watermarking: Fixed a watermark layer that silently failed to write into the metadata of certain real .docx files, and made coverage reporting more accurate.
Recruiter and candidate UX: Added resend-test and extend-time-window actions so recruiters no longer have to delete and recreate assignments (which lost history), fixed race conditions that made the candidate list flash or load stale data, and clarified screen-sharing instructions for candidates.