InterviewLens

A macOS menu-bar app for mock DSA interview practice. It records your screen and mic while you talk through a problem, transcribes and OCRs the session locally, and hands it to the Claude Code CLI for a scored, rubric-based review — one click from "stop recording" to a saved review.md.

Scope: personal use. This is a tool for my own interview prep, not a hardened product. See Known limitations.

---

How it works

Start ─► ffmpeg records screen (10fps)  +  bin/voicerecorder records mic
Stop  ─► mic.wav ─► loudness pass ─► Whisper (local) ─► transcript.txt   [MM:SS] per line
      ─► recording.mov ─► 12 frames spread across the session ─► Tesseract (local) ─► code_ocr.txt
      ─► review_prompt.md (rubric + transcript + OCR) ─► claude --print ─► review.md
      ─► review filed under reviews/ by problem name ─► opened automatically

The review scores four categories out of 10 — explanation quality, speaking efficiency, coding style, correctness — then lists the top three fixes and rewrites your weakest explanation the way a strong candidate would say it. Every point cites an [MM:SS] timestamp you can scrub to in the recording.

What runs where

Step	Runs on
Screen + mic recording	Your Mac
Speech-to-text (Whisper)	Your Mac (model downloads once, then offline)
Code OCR (Tesseract)	Your Mac
Review	Anthropic, via the Claude Code CLI — the transcript, OCR text and screenshots are sent as part of the review

No API keys and no extra cost: the review runs on your existing Claude Pro/Max/Team subscription through Claude Code. The recording video itself stays on disk.

---

Engineering notes

Clean mic audio via a native helper. ffmpeg's avfoundation input grabs the raw mic stream with no noise suppression or gain control. Measured on a MacBook Air: a −33 dB noise floor with peaks clipping at +3 dBFS, which made Whisper produce repetition-loop garbage. native/voicerecorder.m is a small Objective-C helper that records through Apple's Voice Processing I/O audio unit — the path QuickTime and FaceTime use. Result: −60 dB noise floor, peaks at −6.6 dB, no clipping. Things that cost real debugging time:

• VPIO is duplex — enabling only the input bus fails with -10875 and delivers zero buffers; the output bus must be enabled and fed silence.
• AVAudioEngine.setVoiceProcessingEnabled: fails silently in a CLI process, so the helper uses the AudioUnit API directly.
• The built-in mic defaults to 96 kHz while output runs at 48 kHz; the helper pins the input to 48 kHz.

Frames that actually cover the whole session. A fixed fps=1/20 with a 12-frame cap stopped at 3:40 — a 22-minute session was reviewed on its first four minutes, and the last frame Claude saw was an empty function stub. Frame spacing is now derived from the recording's duration, with round=up so the final frame (the finished code) isn't dropped.

Timestamped inputs. The transcript is written as [MM:SS] lines rather than one blob, so the rubric's pacing and dead-air criteria are measured from real gaps instead of guessed.

Resumable pipeline. Every step is skipped when its output already exists, so a review that times out after Whisper's slow pass can be retried without re-transcribing (Reprocess Unfinished Session in the menu).

Devices resolved by name. avfoundation indices shift when hardware connects — plugging in an iPhone (Continuity Camera) pushes "Capture screen 0" from index 1 to 3 — so devices are looked up by name at record time.

---

Setup

./setup.sh

Installs ffmpeg and tesseract via Homebrew, creates a Python virtualenv, installs the Python dependencies, and compiles bin/voicerecorder.

Claude Code CLI (one time):

npm install -g @anthropic-ai/claude-code
claude            # sign in once
claude --print "say hello"

If claude isn't on your PATH, set CLAUDE_CLI_PATH in config.py.

macOS permissions: the first recording prompts for Screen Recording and Microphone access (System Settings → Privacy & Security). You may need to restart the app once after granting them.

To list the capture devices the app will pick:

python devices.py

Running

./run.sh

A ring icon appears in the menu bar. It pulses while recording and becomes a pie-chart progress indicator while processing; the menu's top line names the current phase, e.g. "Transcribing speech… (3/6)".

Configuration

All knobs live in config.py:

Setting	Default	Notes
SCREEN_DEVICE_NAME / AUDIO_DEVICE_NAME	"Capture screen" / your mic	Matched by name; *_INDEX forces an index
USE_NATIVE_VOICE_CAPTURE	True	False falls back to raw ffmpeg mic capture
WHISPER_MODEL	"small"	tiny/base are faster, less accurate on technical terms
MAX_FRAMES_TO_OCR	12	Resolution of the code timeline
FRAME_INTERVAL_SECONDS	20	Minimum spacing (binds only on short sessions)
CLAUDE_CLI_SKIP_PERMISSIONS	True	Unattended writes of review.md; False prompts per new session folder
CLAUDE_CLI_TIMEOUT_SECONDS	480	Raise for long sessions

Optional: to capture system audio too, install BlackHole (brew install blackhole-2ch) and point the audio device at it.

Reviewing past sessions

Each session lives in ~/DSAInterviewPrep/sessions/<timestamp>/ with the recording, transcript, OCR text, frames and review.md. A flat copy of each review is filed by problem name:

~/DSAInterviewPrep/reviews/
├── review-two-sum-2026-08-22_15-20-25.md
├── review-two-sum-2026-08-25_09-10-31.md   <- later attempt, never overwritten
└── review-lru-cache-2026-08-23_11-05-02.md

Claude infers the problem name and writes it as an invisible <!-- question-slug: ... --> marker on the review's first line. Tracking a recurring weakness is then just:

ls ~/DSAInterviewPrep/reviews/review-two-sum-*
grep -l "complexity" ~/DSAInterviewPrep/reviews/*

Open Reviews Folder and Show Last Review in the menu jump straight there.

Known limitations

• OCR on code is imperfect. Tesseract misreads some brackets, indentation and look-alike characters (l/1/I, O/0). The prompt tells Claude to infer intent; for exact code, save a .py/.txt into the session folder before stopping.
• Processing takes real time. Whisper takes seconds to a couple of minutes depending on session length and model size, and the Claude review a few minutes more.
• Claude Code CLI versions differ in whether --print writes files. The pipeline writes review.md from stdout if the CLI didn't. To debug:

  cd ~/DSAInterviewPrep/sessions/<timestamp>
  claude --print "$(cat review_prompt.md)"

• Runs as a script (./run.sh), not a bundled .app.

Stack

Python (rumps), Objective-C (AudioToolbox / AVFoundation), ffmpeg, OpenAI Whisper, Tesseract, Claude Code CLI.