CASE STUDY · DIGITAL AVENUES · 2025–2026
I built the web and mobile clients for Prisium at Digital Avenues: a remote multi-camera live production platform. In 10 months, I shipped a three-workflow platform—with separate paths for live preview and high-resolution recording, a producer studio, and phone-based guest capture.
Remote production needs broadcast-grade multi-camera control over the internet, while guests need to join from their phones. I worked on the web and mobile clients that connect those two sides: a producer choosing what goes live, and participants capturing footage that can also be recorded at high resolution.
I used Socket.IO for signalling and Mediasoup for WebRTC sessions. High-resolution ISO recording stays on a separate capture-and-upload path, coordinated through the web and native clients.
I built the real-time session path with Mediasoup, keeping live preview separate from high-resolution recording.
I built 4–6 participant WebRTC sessions with Mediasoup SFU and Socket.IO signalling. Live preview and high-resolution recording run through separate pipelines.
The producer needs a live view of the session; the recording path needs to preserve high-resolution footage. I separated those responsibilities rather than treating the preview stream as the recording.
I built feed switching, wide shot, screen capture, overlays, LUT grading and program output into the producer studio.
The web client brings source selection and output control into one production surface, alongside the visual tools needed to compose the program.
I orchestrated 4K ISO recording across the web and native Kotlin/Swift code, including Android dual-pipeline capture and resumable chunked uploads to Google Cloud Storage.
Retry and resumable chunks improved recording reliability by 70%. Native capture and web orchestration address different parts of the recording lifecycle.
I built AI camera switching on two platforms using MediaPipe lip detection and audio active-speaker signals.
A rising confidence score prevents flip-flopping. When the signal is unclear, I hold the last speaker; when several people speak, I fall back to a wide shot.
I built the React Native LiveShare app so guests can join through a deep link and broadcast from their phones.
The capture experience includes multi-lens zoom and Bluetooth audio routing, with native capabilities connected to the cross-platform client.
I built a Next.js BFF over FastAPI with Firebase auth, Stripe and SendGrid, reducing integration complexity by 22%.
I delivered three FastAPI workflows using Pub/Sub and Cloud Run for video processing, with Jenkins CI/CD. The BFF connects the web client to those services.
The loudest signal is not always the right moment to cut.
I combined MediaPipe lip detection with audio active-speaker signals. An instant threshold makes the decision too sensitive to a changing signal: small crossings can repeatedly change the selected camera.
I moved to a rising confidence score instead of switching immediately. Sustained evidence earns a switch; unclear evidence holds the last active speaker. When several people speak, the wide shot becomes the fallback rather than choosing one speaker arbitrarily.
A rising score prevents a momentary signal from becoming an immediate cut.
When the signal is unclear, preserve the last active speaker.
When several people speak, fall back to the wide shot.
I’d expand repeatable switching scenarios around ambiguous signals and overlapping speakers. A camera decision needs to make sense even when the evidence does not.
I’d keep capture, upload and recovery as separate failure boundaries. Resumable chunks and retry belong in the design, not as a last-minute fallback.
I’d preserve clear service boundaries while keeping producer-facing behavior predictable. The complexity behind a control should not leak into the decision to use it.