I built the ComplianceQAPipeline, an end‑to‑end AI video compliance auditing pipeline using Whisper, FAISS RAG, Groq LLM, and LangGraph orchestration.
The project started 2026-02, last updated 2026-08.
A live demo is available at https://youtubeaudit.streamlit.app/.
I chose FAISS for vector similarity because it offers low‑latency local search, which fits the pipeline’s local‑first design.
Groq LLM was selected for its fast inference and built‑in strict JSON mode, enabling deterministic structured output.
LangGraph orchestrates the DAG‑style workflow, providing modularity and clear state transitions.
The pipeline is ready for Azure deployment, with optional swaps to Azure OpenAI, Azure Blob Storage, Azure AI Search, and Azure Functions/App Service.
- Key workflow steps:
- Accept YouTube URL.
- Download video locally via yt‑dlp.
- Transcribe audio with Whisper.
- Retrieve compliance rules from FAISS.
- Evaluate transcript with Groq LLM (JSON mode).
- Output structured report with violations, severity, PASS/FAIL, and summary.