Edge A/V Recording & AI Video Analytics Platform
A distributed edge recording system for classrooms, plus a serverless AI layer that turns every recording into an automated quality, compliance, and cost report — built on multimodal LLMs rather than custom-trained computer-vision models.
concurrent classrooms
of A/V data per day
Classrooms needed reliable audio and video recording on Wi-Fi-only networks, where power cuts and dropped connections are normal. A lost recording cannot be re-shot.
Once recorded, every session also needed review: is the camera framed correctly, who is in the room, are staff in uniform, and how good was the lesson? Checking hundreds of gigabytes of footage a day by hand does not scale.
Raspberry Pi devices with a camera and wireless mic record each classroom and upload to Amazon S3 through retryable, fault-tolerant upload pipelines.
An S3-triggered AWS Lambda compresses each recording with FFmpeg (24fps, 480p, H.264, CRF 30), skips files already marked compressed to avoid loops, and explicitly invokes the analysis Lambda.
The containerized analysis Lambda (Python 3.12, 2–4 GB RAM, 900s timeout) claims each job in DynamoDB, so a recording is never analyzed twice even if invoked more than once.
OpenCV samples two frame sets per video for two GPT-4.1 Vision passes. In parallel, the audio track is chunked, transcribed with Groq-hosted Whisper, and scored by an LLM against a weighted rubric.
Results are written to S3 as one JSON report, then sent to the FastAPI backend through an HMAC-SHA256-signed callback and stored in MySQL, with retries, human-review overrides, and versioned pipeline configuration.
Visual checks run as GPT-4.1 Vision passes over sampled frames: one audits camera health (misaimed, cropped, blank, or distorted framing), the other counts trainers, associates, parents, and children and checks uniforms.
Trade-offNo detector to train, label data for, or maintain — new checks and custom labels are a configuration change, not a new model.
A DynamoDB claim guards every analysis job, and the compression step checks a compressed flag before running, so retries and duplicate invocations never double-process a recording.
Trade-offOne extra write per job buys safe retries everywhere in an event-driven pipeline.
The LLM rates each dimension of a transcribed lesson; the final quality score is a weighted formula in code, so the model never self-scores.
The pipeline hands results to the application through an HMAC-SHA256-signed callback. Reviewers can override AI results, and pipeline configuration is versioned.
Trade-offAI output is treated as a first pass, not a final verdict.
concurrent classrooms recorded over Wi-Fi-only networks
of audio and video processed per day
near-zero data loss across all edge devices
Every recording now arrives with an automated quality, compliance, and cost report, with per-call LLM cost tracked across every vision, transcription, and scoring call.