MszaAI

Turning church livestreams into searchable transcripts, sermon studies, and source-grounded answers—so valuable homilies and parish information never disappear.

Comment

Inspiration

Every day, valuable homilies and parish information disappear inside hours of livestream video. Some churches do not provide recording playback at all: the broadcast simply ends and the content is gone. This especially affects people who missed the Mass, are homebound, or want to revisit a sermon later.

We built msza.ai to turn temporary church livestreams into lasting, searchable and verifiable public knowledge.

What it does

msza.ai monitors more than 200 church livestream sources and uses local audio checks to detect when a Mass is likely beginning. Only then does it activate transcription, keeping unnecessary processing and API costs under control.

After a broadcast, users can:

  • read the transcript and a structured timeline;
  • find sermons, announcements and summaries;
  • explore daily, weekly and monthly sermon studies;
  • search the public archive;
  • follow selected churches and receive notifications;
  • translate published content on demand;
  • ask an AI assistant questions across the archive.

The public AI assistant retrieves relevant passages from real homilies and studies. Its answers include clickable citations, allowing users to inspect the original sermon instead of trusting an unsupported AI response.

What we added during OpenAI Build Week

msza.ai existed before the submission period. During Build Week, we meaningfully extended it with GPT-5.6 and Codex by building and refining:

  • a public AI assistant for the complete archive;
  • source-grounded retrieval across sermons and studies;
  • clickable citations leading to the original passages;
  • transparent explanations of the assistant's limitations;
  • stronger privacy and cost controls;
  • an English user-facing interface with shareable language links;
  • on-demand translation while preserving original source content.

The submission focuses on these additions and how they transformed the existing archive into an accessible, verifiable knowledge system.

How we built it

The frontend uses React, TypeScript and Vite. The Node.js and Express backend communicates with the interface through HTTP and WebSockets, while SQLite stores cameras, sessions, transcripts, studies, citations and user preferences.

yt-dlp and FFmpeg handle livestream audio. Cheap local checks run while a camera is idle, and transcription services are activated only after the state machine detects a likely live event. This prevents the system from sending hundreds of silent livestreams to paid APIs.

The AI assistant operates over a read-only public corpus. Retrieval identifies relevant transcript and study passages, while the answer layer preserves source metadata so every supported claim can link back to its evidence.

How we used GPT-5.6 and Codex

GPT-5.6 and Codex helped us work across the entire product instead of treating each feature in isolation. We used them to trace data through the React frontend, Express backend, retrieval layer and SQLite storage; implement the public assistant and citation flow; improve the bilingual UI; identify privacy and cost risks; test edge cases; and keep documentation aligned with the implementation.

Codex was especially valuable when changes crossed several layers at once. It helped us preserve existing behavior while extending a live production application under a very short deadline.

Challenges we faced

Church livestreams are unpredictable. Streams change URLs, remain silent for long periods, contain music before speech and sometimes disappear without an archive. Distinguishing a real Mass from unrelated audio without continuously paying for transcription required a careful local-first state machine.

A second challenge was trust. An AI answer about a sermon should never appear more certain than its sources. We therefore designed the assistant around retrieval, citations and explicit limitations rather than unrestricted generation.

We also needed to make an originally Polish product understandable to international judges without automatically altering or mistranslating the underlying religious source material.

Accomplishments we are proud of

We turned temporary broadcasts into searchable public memory while keeping source verification one click away. The same workflow now connects livestream detection, transcription, structured archives, sermon studies, notifications and grounded AI answers.

Most importantly, people who could not attend or watch a broadcast live can still discover what was said and return to it later.

What we learned

Reliability and trust matter more than adding another generic AI chat box. Local checks can dramatically reduce cost, retrieval makes answers more useful, and visible citations let users decide whether an AI interpretation is justified.

We also learned how effective GPT-5.6 and Codex can be as engineering collaborators when they are given the product context, constraints and responsibility for following changes across the complete system.

What's next

We plan to improve automatic event detection, expand international accessibility, strengthen retrieval evaluation and make following churches and receiving notifications even easier across web and mobile.

Built With

Share this project:

Updates