Problem
Anonymous video needs coordination and privacy before either participant chooses to reveal.
Constraints
Anonymous sessions, scheduled admission, mutual reveal and media-level privacy were explicit requirements.
What I did
SkipChat is a learning project: a live video product I built around a strict privacy contract. Anonymous adults can talk first, stay visually protected at the start of a match, and reveal only after both participants choose to.
Product constraints
The system uses scheduled live windows to concentrate demand. Admission is scheduled, but an admitted user can continue through multiple 1:1 matches after the window closes.
Constraint | Implementation |
|---|---|
No account required | Anonymous session and short-lived admission. |
Scheduled liquidity | Availability is checked before queue admission. |
Continuous session | Skip and partner leave return the user to matching. |
Mutual reveal | One coordinator resolves both choices. |
Real privacy | A blurred media track is sent before reveal. |
Abuse handling | Reports, bans, and audit records persist in D1. |
State ownership
Durable Objects own queue membership, match lifecycle, timers, reveal decisions, WebRTC signaling, and socket cleanup. The Coordinator is the control plane; WebRTC remains the peer-to-peer media plane.
The browser first creates an anonymous session and requests a one-use queue admission. The Coordinator rejects expired, reused, or session-mismatched admissions before accepting queue.join. This keeps eligibility policy on the HTTP path and reserves the WebSocket protocol for live state.
The matcher is intentionally simple. It finds a waiting participant who is not the user’s previous partner, creates a match and round, persists the relationship, and notifies both sockets. The hard problems are lifecycle correctness, media privacy, and rematching—not recommendation quality.
Media-level privacy
A CSS blur would be false privacy: the browser would still receive a clear remote track and merely paint it differently. SkipChat renders the camera into a blurred canvas, sends that derived track through WebRTC, and replaces it with the raw camera track only after the Coordinator emits reveal.mutual.
The guarantee therefore exists in the media graph, below the React view. Before mutual reveal, the clear outbound video track is not present in the peer connection.
Operational model
Live state belongs near the live sockets. D1 stores durable records such as anonymous sessions, reports, bans, and availability rules. Analytics capture admission, matching, reveal, media, and rematch lifecycle events.
That separation matters because an HTTP endpoint can be healthy while the product loop is down: nobody may be matched, reveal may never resolve, or rematching may be too slow. Product availability is measured through lifecycle transitions, not only request success.
SkipChat is a compact example of cross-layer product correctness. A reveal button depends on a Durable Object decision, a persisted match state, a WebSocket event, a peer connection, and a video-track replacement. The state machine has to be correct before the interface can feel simple.
Result
A live video product with Durable Object coordination, WebRTC media and persistent abuse-handling records.
Tradeoffs
Scheduled windows concentrate demand; coordination and media transport stay separate. A CSS blur cannot supply the same privacy contract as a blurred outbound track.