/ Works

/ 2026

/ GBI Baranangsiang Evening Church

GBIBECChurchOperationsPlatform

Live sermon transcription that summarises itself, an automatic summary-to-post pipeline, an Instagram sync, and an AI poster generator that redraws its own images — behind a portal with seven public forms and a 22-page admin.

/ Client

GBI Baranangsiang Evening Church

/ Role

Solo — architecture and build

/ Timeline

Mar — Aug 2026

/ Stack

Next.js / TypeScript / Pinecone / Firebase / Google APIs / TipTap / pptxgenjs / Docker / Python

GBI BEC — Church Operations Platform

About the project

/ Overview

A church operations platform: a public portal with seven forms, a 22-page admin behind it, four containerised sync services, and an AI layer that transcribes sermons, summarises them, and publishes the result. Everything a volunteer team would otherwise do by hand on a Sunday evening.

/ The Challenge

The work is weekly and the people are volunteers. A sermon has to be captured, transcribed, summarised, turned into a post and pushed to the site and to Instagram; a service needs its songs found, its lyrics fetched and its slides built; forms need collecting and reporting. Done manually that is hours every week, it slips the moment someone is away, and the congregation's questions go unanswered because nobody has time to answer them twice.

/ The Approach

Each of those became a pipeline rather than a task. Sermon capture starts with the stream and stopping it triggers the summary, which can be regenerated, combined with manual notes, or handed to a notetaker — then a publish chain pushes it to the news section without retyping. Posters and post copy are generated and the images adjusted in place. A Pinecone-backed RAG chat answers congregation questions and logs every question it could not answer, so the gaps are visible rather than lost. The song library fetches lyrics, searches the web and segments them into verses, then exports service slides to PowerPoint. The syncs — YouTube live, YouTube sermons, Instagram — run as Cloud Run Jobs on Cloud Scheduler, and the embedding service runs ONNX Runtime rather than PyTorch because it is roughly a tenth of the image size.

/ The Result

One portal and four services now carry the weekly cycle. Over sixty API routes sit behind seven public forms and a 22-page admin, and the parts that used to be somebody's evening — transcribe, summarise, post, sync — happen on a schedule whether anyone is available or not.

/ What it took

  • Live sermon transcription: capture starts with the stream, and stopping it triggers the summary — regenerate, combine with manual notes, or notify the notetaker from the same record
  • Summary-to-post pipeline: a finished sermon runs a publish chain straight out to the news section, no retyping
  • AI poster generation that adjusts its own images, and AI-written post content to go with it
  • Instagram sync as its own containerised service, alongside YouTube live and YouTube sermon syncs
  • A Pinecone-backed RAG chat answering congregation questions, with a chat-miss log recording everything it could not answer
  • Service sets assembled from a song library that fetches lyrics, searches the web and segments them into verses — then exports to PowerPoint via pptxgenjs
  • Seven public forms with reporting behind a 22-page admin — over sixty API routes in total

/ Results and outcomes

API routes
60+
Admin pages
22
Cloud Run services
4