For pastors and evangelists
Poimens keeps your scholarship and your sermons in two separate libraries — one for facts, one for cadence — then drafts a brief, an outline and a full manuscript that sounds like you. Every draft is scored for scripture, quotations and originality before you preach it.
Self-hosted. Your manuscripts stay on your own server and are never used to train a model.
Two libraries
Blending your commentaries with your old sermons is what makes generated preaching sound generic — or confidently cite something you never said. Poimens separates them when you import, and keeps them separate on every search.
Doctoral work, commentaries, theology, history and misc documents. Chunked and embedded so every historical, exegetical or theological claim can be traced back to a source you supplied.
Transcripts of sermons you have actually preached. Analysed for cadence, sentence rhythm, how you introduce Scripture, how you handle application and how you land a conclusion — then distilled into a reusable voice profile.
How it works
Nothing is generated in one opaque leap. You review the research, the brief and the outline before a single paragraph of manuscript is written.
Drop in your research and your sermon transcripts. Text is extracted, split on real structure and embedded locally.
Poimens reads across your sermons and writes down how you actually preach — rhythm, transitions, vocabulary, application habits.
Give it a passage, a theme and a length. It retrieves the evidence, then produces a brief, an outline, and a full manuscript section by section.
Every draft is scored: do the references exist, can the quotations be traced, does it reuse anything you already preached, is it long enough.
Before you preach it
A language model will happily invent a quotation, misnumber a verse, or hand back something you already preached in 2023. Poimens checks for all three and shows you exactly where to look.
Needs review
Up to 30 consecutive words identical, 35% of the passage's phrasing shared. Rewrite these in fresh language.
Confirm each one before preaching — a quotation that cannot be traced to your sources or the Bible text is the classic failure mode.
Within the same manuscript.
Everything in the studio
Semantic search for ideas, keyword search for exact references, fused so a query about groaning and a query about “Romans 8:22” both land.
Cadence, sentence rhythm, question rate, direct address, characteristic vocabulary and transition habits — measured and written down.
Three reviewable stages. The manuscript is composed section by section so it reaches the length you asked for, not a 500-word summary.
Scripture validity, quotation traceability, originality against your own past sermons, repetition, and outline drift.
Passage reading, full-text search and concordance across four public-domain translations — KJV, ASV, Webster and Basic English.
Plan a series arc with a real passage sequence, then schedule it on a preaching calendar with status from planned to preached.
Private by default
Embeddings run locally on your own machine, so your documents are never uploaded to be indexed. Choose which model writes for you — an offline model that never leaves the building, or a hosted one for stronger prose. Either way only the assembled prompt is sent, never your library.
openrouter/deepseek/deepseek-chat-v3.1Create a workspace, import a few commentaries and two or three transcripts, and generate a draft you would actually be willing to preach.
Next 60 days from your calendar.
Where each project sits in the pipeline.
Generation and embedding endpoints reported by the server.
Drop a whole sermon archive or a folder of papers at once. Extraction and embedding run in the background.
Doctoral theses, commentaries and academic papers. These supply the facts, definitions and citations the model is allowed to lean on.
Your own transcripts and manuscripts. These teach the model how you preach — sentence rhythm, transitions, direct address.
Dense + sparse retrieval fused with reciprocal rank. Scores and both ranks are shown so you can judge a hit.
Built only from the Sermons corpus. This is what makes a draft sound like you rather than like a model.
A voice profile is distilled from transcripts and manuscripts in the Sermons corpus. Import at least one sermon, wait for it to finish indexing, then rebuild.
Computed from your indexed sermons — no model involved.
Plain-text guidance injected into every generation prompt.
Offline scripture. Imports stream verse text into the local database once.
Book, chapter and optional verse range.
Full-text search inside one translation.
Every occurrence of a single word.
Wording-similarity suggestions — not a curated cross-reference set.
Imported chapters and their verse counts — see at a glance what is loaded.
—
A title and dates are enough; the AI planner can draft the arc.
Click any day to add an entry. Click an entry to edit its status or delete it.