Interview Question Recall Engine

Delivery approach

1. S

   Situation

   Candidates often left an interview remembering only fragments of the questions. A full recording or raw transcript took too long to review, and ordinary transcription mixed speaker errors, filler, repeated prompts, and the candidate's own questions. The recordings also contained personal conversation, so failed uploads, temporary data, and deletion needed explicit handling.

2. T

   Task

   Turn one recording into an ordered list of interviewer questions with timestamps and enough context for later preparation. Processing had to tolerate noisy audio and overlapping speakers, mark uncertainty instead of inventing a question, resume after interruption, and keep the interface focused on recall rather than transcript editing.

3. A

   Action

   An asynchronous pipeline validates the upload, splits long audio, transcribes each chunk, aligns timestamps, and reconciles speaker labels before extraction. A separate classification pass identifies interviewer turns and distinguishes questions from greetings, explanations, and acknowledgements. It merges adjacent fragments, collapses near-duplicates, and keeps the original timestamp on every extracted question. Ambiguous speakers are marked as uncertain rather than assigned silently. Checkpoints allow one failed chunk to be retried without transcribing the whole file again. Completed and abandoned files follow a defined retention and deletion path. Manually labeled conversations test accents, background noise, interruptions, multi-part questions, and speaker changes. Transcription usefulness and question precision are measured separately.

4. R

   Result

   One upload returns a reviewable question list instead of a full transcript. Timestamps let the user check uncertain output against the recording, and chunk-level retries prevent a single failure from restarting a long file. The product stays focused on question recall rather than adding meeting-summary features.