AI Operator: How Kerygma Follows a Sermon in Real Time

A while ago, I shared that I was building a recursive Bayesian belief filter for real-time autonomous decision-making. I did not explain what it was for.

This is what I was building.

AI Operator beta is now available in Kerygma v1.1.4.

With one switch, Kerygma can listen to a sermon, identify the scripture being discussed, and display the relevant passage without someone constantly operating the keyboard.

It can follow a preacher verse by verse, advance through longer passages screen by screen, and recognise scriptural quotations even when the preacher never mentions the book, chapter, or verse.

Building it was not simply a speech-recognition problem. It was a decision-making problem.

Why simple rules fail

The obvious way to build something like this is a rules engine

e.g. If the transcript contains “John 3:16,” display John 3:16.

That works in a controlled demonstration, but it can quickly fall apart during a real church service.

Live sermon audio is noisy and unpredictable. Transcription systems mishear words. Preachers paraphrase verses, pause halfway through references, move into stories, and return to the passage several minutes later.

Ordinary words can also resemble Bible-book names. A system might hear “is” and assume Isaiah, or interpret “am” as Amos.

Sometimes, the preacher quotes an entire verse without stating its reference at all.

A basic rules engine reacts immediately to each possible match. During early testing, Kerygma once put dozens of wrong verses on a screen in a live service. Every false trigger goes straight to the congregation.

We know in a live environment, being fast and wrong is worse than being slightly slower and right.

From triggers to belief

AI Operator works differently.

Instead of reacting to individual moments, it holds a running belief about what's happening in the sermon, and it updates that belief with every new piece of evidence.

This is where recursive Bayesian filtering comes in.

In simple terms, the system repeatedly asks:

Based on everything I have heard so far, which passage is most likely being preached, and how confident am I?

Each new sentence provides evidence.

A clearly spoken reference provides strong evidence. A close quotation may provide equally strong evidence. A loose paraphrase provides weaker evidence, while an isolated and ambiguous word provides almost none.

The system combines that new evidence with what it already believes about the sermon.

It does not display a passage until the confidence is high enough.

Because of this, one uncertain word cannot suddenly send Isaiah or Amos to the screen. The system waits for supporting evidence before making a visible decision.

Does it feel safe to use 

While building this feature, accuracy was only part of the problem. The system also needed to behave calmly.

A reactive system flickers. It shows a verse, doubts itself, swaps to another, jumps back. On a screen in front of a congregation, that reads as chaos.

That uncertainty may be acceptable inside a debug console. It is distracting when projected in front of a congregation.

AI Operator avoids that by building its decisions from accumulated evidence rather than isolated moments.

Once a verse is live, the system requires genuinely stronger evidence before switching to something else. A brief phrase or weak match should not be enough to interrupt the passage already being displayed.
It also keeps track of movement within a passage.

When the preacher is reading through a long passage,AI Operator progresses forward through the verses and presentation screens at the pace of the sermon. It avoids jumping backwards simply because an earlier phrase is repeated.

The goal is not to behave like autocomplete.

The goal is to behave like an experienced media operator:

Listen. Wait. Confirm. Commit. Advance.

The part I am most proud of

The hardest case is a scripture quotation with no spoken reference.

A preacher may say:

“Whatever a man sows, that shall he also reap.”

They may never say “Galatians 6:7.”

AI Operator can recognise the quotation, even when the wording is slightly different, and match it to the most likely passage.
The quotation is treated like every other piece of evidence. A close match may be strong enough on its own. A looser match can be combined with the current sermon context, previous references, and nearby transcript content.
When the system finds a likely passage, it can bring the scripture up and show the operator what it detected.
Watching AI Operator identify verses that I had not even realised the preacher was quoting has been one of the most satisfying parts of building Kerygma.

It does more than identify a verse

Recognising scripture is only one part of the problem.

AI Operator also has to decide what should happen next.

Should it display the passage immediately?

Should it wait for more evidence?

Is the preacher still discussing the current verse?

Have they moved to the next verse in the passage?

Is a newly detected reference the main scripture or only a passing mention?

These are not simple lookup questions. They are decisions made over time.

That is why AI Operator combines multiple parts of the system:

• Ears: listens to and transcribes the sermon.
• Citation: detects spoken references and scripture quotations.
• Context: understands the current passage and recent sermon activity.
• Memory: keeps track of previous evidence and decisions.
• Following: monitors the preacher’s progress through the passage.
• Director: decides when to wait, display, advance, switch, or recover.

Each part contributes information, but Director is responsible for the final decision.

Why context matters

Imagine that the preacher is teaching from John 3 and briefly mentions Genesis.

A basic system may immediately replace John with Genesis because it detected a new Bible reference.

AI Operator asks a different question:

Is Genesis now the main passage, or was it only mentioned to support the current point?

The answer depends on context.

If the preacher mentions Genesis once and immediately continues reading from John, the system should remain in John.

If the preacher begins quoting Genesis, discussing its verses, and returning to it repeatedly, the system’s belief gradually shifts.

This prevents the screen from jumping to every passing reference mentioned during the sermon.

It also allows the system to recover when the preacher genuinely changes direction.

Autonomous does not mean invisible

AI Operator is designed to reduce the workload of the media team, not hide what the system is doing.

The operator can still see the selected passage, understand why it was chosen, and intervene when necessary.

This matters because church services are live, human, and unpredictable. Good automation should not remove control. It should reduce repetitive work while keeping people informed.

Try AI Operator

AI Operator is currently in beta in Kerygma v1.1.4. You can enable it using the switch beside the Translation menu.

It is already being tested during real church services, but there is still more to improve: scripture recognition, pacing, confidence calibration, decision explanations, and recovery when the sermon changes direction unexpectedly.

Real-world feedback is essential.

If AI Operator makes an unusual decision during your service, I want to hear about it; Send it to support@kerygma.dev. Every sermon provides new examples that can help make the system more accurate and dependable.

If you work in church media—or you are curious about what this algorithm looks like when applied to a live sermon—try AI Operator and tell me what it gets right, what it misses, and what you would like it to do next.

Visit Kerygma at kerygma.dev

— David Ukonu, Founder of Kerygma