Inspiration
Dating apps are tired. People swipe through photos without talking, match without meaning it, and ghost without consequence. The last time strangers talked on the phone to meet someone was 1-800 dating lines and 900-number voice personals in the late 90s and early 2000s. The mechanism worked; the technology died.
DeadLine brings that mechanism back with modern anonymity. Two weeks of evening calls at 8-10 p.m. local, two scheduled calls per night, all anonymous voices. Every identifying detail that slips out (a name, a phone number, an email, an address) turns to line static before the other person hears it. Hearing a returning caller is how both discover that continuing was mutual. No match notifications, no rejection counts, no profile browsing. Chemistry develops through conversation.
What it does
A cohort of 32 people (16 x 16) is pre-matched on preferences and dealbreakers before anyone is admitted. Nobody joins the cohort until at least four reciprocally-eligible introductions exist for them. Two 45-minute calls per night, from 8:00-8:45 and 9:00-9:45. Short private confessional between calls. Private selections stay hidden; a returning caller's voice is the only continuation signal. Declined calls get a brief answered-line experience (quiet line noise then closure) so nobody is left waiting. Technical failure never masquerades as rejection: the four states (misconduct, missed, tech fault, no chemistry) stay distinct in both UX and the reporting system.
Identity is verified by the operator before the first date. Callers remain anonymous to each other. "Hang up and report" with human review is private and distinguishes misconduct from technical failure.
How I built it
I built DeadLine in Kiro, from requirements and architecture through application code, tests, and debugging. Concept work started in ChatGPT; the build itself is Kiro-native.
Spec-driven development came first. The spec at .kiro/specs/anonymous-call/ defined the narrow first milestone: two people could join a call, neither could access the other's raw microphone audio, and a processing failure would hold the line instead of exposing the original audio.
Kiro's research changed the architecture when it found the first track-permission approach was insufficient. A modified caller client could change per-track subscription permissions and undermine the restriction. Kiro revised the design to split audio intake from playback: callers publish into an ingest room where signed permissions prohibit subscribing, listen in a separate playout room where they cannot publish, and the worker bridges the two and controls its output tracks.
The redaction pipeline uses a bounded audio buffer, transcript detection via a direct Deepgram adapter, and sample-level replacement with telephone-band static on identifier spans. Deterministic transcript fixtures let me test chunk boundaries, revised recognition results, partial identifiers, and delayed processing before any external speech-recognition call. End-of-speech handling respects the processed interval so release never exposes speech arriving after the confirmed result. Provisional masking fires for incomplete number sequences.
Isolation tests verify what the listener actually receives: one caller publishes a 440 Hz tone and the other publishes 660 Hz; received audio is analyzed to confirm each caller hears only the partner's tone. Tests also exercise manual subscription attempts, reconnects, worker restarts, output-track cleanup, and the duplicate-reader case when callers are in the room before the worker starts.
Operator presence checks and one-second heartbeats detect worker absence. The usage budget is shared across callers and persists across restarts. The five-minute test timer starts with the first stream, not at worker boot. MIP opt-out is sent on every connection.
Challenges I ran into
The first manual browser attempt stayed on "Ringing" even though component tests passed. Kiro traced the connection and integration path and replaced the Deepgram plugin with a direct adapter. That unblocked the full ingest-worker-playout flow.
A repeated output-failure/restart glitch occurred after an abruptly killed test. It did not recur in later runs and diagnostics improved, but the root cause remains unexplained. Finding the mechanism before live-user testing is on the list.
The real-speech listening test is still pending as of this draft. Mock-recognizer tests prove the pipeline behaves correctly on supplied transcripts. They do not establish that a real recognizer will understand speech reliably enough to redact actual names, phone numbers, or email addresses. That test runs separately with headphones and the identifier list before anything else lands.
Accomplishments I'm proud of
- Kiro caught a security flaw in the first architecture approach (track-permission bypass) and revised the design to the ingest-playout split before it shipped
- Measured relay isolation with real tone-tests, not just connection checks
- End-of-speech handling corrected to respect the processed interval after review exposed the leak case
- 18 simulated redaction tests, 5 mock adapter tests, 10 local isolation tests, and 6 mock-backed browser checks passing
- Honest limits documented throughout; mock recognition is not presented as proof of real-speech redaction
What I learned
The hardest part of this build is the listener's ear. Proving that identifying details actually turn to static, in real speech conditions, before the other side hears them, is the work that gates everything else. Every architectural decision downstream of that gate reads easier once the gate is proven.
Spec-driven development with Kiro made the tradeoffs legible early. The specs documented what had to happen before audio could be released, what should happen when processing fell behind, and which claims the project could actually support. When Kiro's research contradicted the first design, the specs made the architectural revision a clean handoff.
What's next for DeadLine
- Real-speech listening test with headphones against the full identifier list (names, phone numbers, emails, addresses, split identifiers, harmless numbers)
- Root-cause the output-failure glitch before live-user testing
- Identity verification vendor selection and integration
- Misconduct review process definition before cohort recruitment
- First cohort recruitment at MTSU and in Nashville, positioned as a product for people who want conversation-first dating without applying to a reality show
- Podcast and stage-reveal production as a growth channel after the dating product stabilizes
