The administrative burden of supervising engineering capstones is notoriously high, but Cyrus Wong, a Senior Lecturer at HKIIT and Google Developer Expert, has built a pipeline that uses Gemini 4 to cut through the noise. His latest project, detailed in Part 8 of his Google AI Classroom series, tackles the specific pain points of academic supervision: students submitting vague or AI-generated fake logs and the chaos of multilingual, multi-person lab consultations.

Auditing the 'ChatGPT Logbook' Epidemic

Wong’s system addresses a common failure mode in modern education: students pasting LLM-generated buzzword salads into weekly progress forms. The pipeline uses a pre-meeting AI auditor that cross-examines a student’s current log against their prior weeks’ history. If a student who wrote 'done' for three consecutive weeks suddenly claims to have implemented 'real-time microservice telemetry' and 'RBAC authentication,' Gemini 4 flags the entry as 'suspected_ai_fake' and generates targeted code-walkthrough questions to expose the bluff during the live meeting.

Mobile-First Multimodal Capture

Recognizing that traditional screen-sharing is inefficient for hardware-heavy projects, Wong replaced PC screen capture with a single-device mobile studio. The supervisor uses one smartphone to record continuous voice, snap 0.1-second photos of whiteboard diagrams or IoT breadboards, and auto-slice 720p videos of live code demos. This approach eliminates the friction of waiting for multiple students to authenticate screen shares, allowing the supervisor to capture physical artifacts and digital outputs seamlessly without pausing the consultation.

Two-Stage Diarization with Resilience

The core technical innovation lies in the two-stage analysis pipeline. Stage 1 uses gemini-3.5-transcribe-preview to handle acoustic diarization, with automatic fallbacks to gemini-3.8-flash if the model truncates transcripts due to long silences or low-volume speech. Stage 2 then feeds the diarized text, alongside the raw 720p videos and photos, into gemini-4-argon for multimodal verification. This ensures that the final report accurately attributes technical contributions to specific students, even when they are code-switching between Cantonese, Mandarin, and English in a noisy lab environment.

Key Takeaways

  • Gemini 4 is used not just for generation, but for critical auditing of student work against historical patterns.
  • The pipeline handles noisy, multilingual audio by using dedicated transcription models with self-healing truncation recovery.
  • Mobile-first capture (720p video + instant photos) outperforms traditional screen-sharing for hardware and whiteboard-based engineering projects.

The Bottom Line

This isn't just about saving time; it's about enforcing academic integrity through rigorous, automated cross-examination. Wong’s architecture proves that LLMs are most powerful when they act as skeptical auditors rather than passive generators.