The Problem
3–8 sec
the gap when a pastor calls an unplanned verse mid-sermon
Church media operators run two jobs at once: advancing planned slides and hunting for surprise references. When the pastor improvises, something gets dropped.
What Was Built
Listen, detect, display — under two seconds.
- Auto-detect mode — mic listens continuously; a custom detection engine recognises all 66 Bible books by spoken name, abbreviation, and number words
- Cue list mode — build a service plan in the Builder tab, arm it, and advance with the Space bar
- Projector output — verse appears in the display panel and a full-screen window instantly; operator never loses focus
- Custom text — push lyrics, announcements, or blank slides at any moment during the service
- Custom test harness — 80+ adversarial test cases across 12 categories including false-positive traps, number words, verse ranges, and mid-sentence detection; results generated live in the browser
Try It
Type any spoken Bible reference and watch the engine detect it.
Waiting for input…
Accuracy
The detection engine was tested against 67 spoken variations before deployment.
Live results — computed on this page
The engine runs inline here. Every figure below is generated by actually executing the detection logic against the test suite right now — not hardcoded.
—
accuracy across all spoken variations
The test suite running in the browser — 67 spoken variations, 0 failures. Includes adversarial cases designed to expose false positives.
The Decision
Replaced a paid AI API with a local regex engine. Zero running cost.
- Local detection handles all 66 Bible books by name, abbreviation, and number words — no subscription, no latency, no credits burning
- Number-word normalisation converts spoken phrases like “one hundred and nineteen” to 119 before matching — every chapter and verse in the Bible, spoken naturally
- Ambiguous-word filter suppresses false positives on common English words that double as book names — “job”, “mark”, “acts”, “numbers”, “ruth” only fire when accompanied by “chapter”, “verse”, or a colon
- “And” as verse separator — “Philippians four and thirteen” now resolves to Philippians 4:13, not just chapter 4
Outcome
Deployed. Free for any church. Detection runs entirely in the browser.
- No backend, no server costs, no API key — any church can open the URL and use it today
- Custom NLP without a model — the regex engine is faster and cheaper than any AI API for this specific matching task, and verifiably correct
- Self-proving portfolio piece — the accuracy chart on this page runs the actual test suite on load; the number is never hardcoded
- Demonstrates knowing when not to use AI — a deliberate constraint, not a gap; adversarial testing exposed three failure modes that were fixed before release