The Process

From source document to published post in under 60 seconds

Eight steps. Three Claude AI calls. Every claim individually verified against your source. Here's exactly what happens when you generate content with Zeplyn.

Step 01

Upload your source document

Start with any scientific source material. Zeplyn accepts PDF research papers and whitepapers, DOCX files, public URLs, and plain text paste. There's no formatting required. Upload the raw document as-is.

  • PDF: research papers, whitepapers, clinical briefs, regulatory filings, datasheets
  • DOCX: internal reports, technical documentation, literature reviews
  • URL: public journal articles, press releases, product pages
  • Plain text: copy-paste from any source
🔒 File is deleted from Zeplyn servers immediately after extraction. Never stored.
Upload Source
📄

Drop your PDF or DOCX here

or paste a URL / text below

📑

deNOVO_HCP_whitepaper.pdf

2.4 MB · Uploading…

Step 02

AI extracts and structures the facts

Zeplyn's extraction layer reads your entire document and compresses it into a structured fact set: key claims, quantified data points, experimental outcomes, product specifications, and named entities, preserving the scientific meaning without the filler.

  • Numerical data preserved with units and context
  • Named compounds, assays, and protocols identified
  • Statistical outcomes and confidence levels captured
  • Source location indexed for every extracted fact
Extracted Facts

Fact Set: deNOVO_HCP_whitepaper.pdf

DATA HCP assay sensitivity: 99.2% (p<0.001) · Source: Table 2, p.4
RANGE Bioreactor scale: 10L to 200L validated · Source: Methods, p.6
SPEC Detection limit: 0.5 ng/mg product · Source: Results, p.7
CLAIM ICH Q6B purity threshold: met across all scale points · Source: Discussion, p.9

Step 03: AI Call 1

Completeness check, before a word is written

The first of three Claude AI calls runs before content generation begins. It reviews the extracted fact set and asks: is there enough here to write accurate, complete content? Or are there gaps that could force the generator to invent?

  • Missing endpoints or outcome measures flagged
  • Undefined abbreviations caught early
  • Insufficient context for regulatory claims identified
  • You can upload a supplementary source to fill gaps before proceeding
Completeness Check: Call 1
✓ 4 facts extracted ⚠️ 1 gap found
⚠️ Gap detected: Mechanism of action for HCP removal is described qualitatively but lacks quantified step-efficiency data. Content about the purification process may require inference.

Suggestions:

Upload purification step-efficiency table (optional)
Proceed anyway. The gap will be flagged at grounding stage

Step 04

Select your format and set your tone

Choose from 11 content types, from a 280-character X thread to a 1,800-word blog post. Then answer two guided questions to calibrate the output tone. No prompt engineering needed.

  • 11 content types with platform-specific formatting rules
  • Audience: Healthcare Professionals, Scientists, Investors, General
  • Goal: Build Credibility, Generate Awareness, Drive Enquiries, Educate
  • Brand Kit and Voice Profile applied automatically
Format & Tone

Content Type

✓ LinkedIn Brand Post
Blog Post
X Thread
Substack

Audience

Healthcare Professionals Scientists Investors

Goal

Build Credibility Drive Enquiries

Step 05: AI Call 2

Content generated from your facts only

The generation call uses a tightly constrained context window: only the verified facts extracted from your document, your brand kit, and your tone settings. There is no internet access, no knowledge base fallback, and no room for fabrication.

  • Constrained to your document's verified fact set
  • Applies your Voice Profile and Brand Kit automatically
  • Respects platform character limits and formatting conventions
  • Marks anything it cannot directly substantiate for the grounding audit
Generating: Call 2

LinkedIn Brand Post: NovaCDMO

Bioreactor scale-up shouldn't mean accuracy trade-offs.

Our latest HCP assay validation confirms 99.2% sensitivity across the full manufacturing range, from 10L development scale to 200L commercial production, with a detection limit of 0.5 ng/mg product.

ICH Q6B thresholds met at every stage. No blind spots between Phase I and commercial launch. 🔬

Generating…

Step 06: AI Call 3

Grounding audit: every claim, individually checked

After generation, the grounding audit runs. A dedicated Claude call reads the output sentence by sentence, cross-referencing each factual claim against the original source document. Every claim gets a tag.

  • ✅ Grounded: directly supported by source text or data
  • ⚠️ Unverified: plausible but not explicitly stated in source
  • ❌ Contradicts Source: factually inconsistent with source
  • Overall grounding score (%) shown per output
  • Source page reference shown for each grounded claim
Grounding Audit: Call 3

Claim Verification

94% Grounded
✅ Grounded

"99.2% sensitivity": confirmed Table 2, p.4

✅ Grounded

"10L to 200L": confirmed Methods, p.6

✅ Grounded

"0.5 ng/mg detection limit": confirmed Results, p.7

⚠️ Unverified

"No blind spots": inferred from data, not explicitly stated

Step 07: Optional

Generate a branded image

Optionally generate a supporting image using GPT Image 2. Describe the visual, or let Zeplyn suggest one based on your content. Your logo is composited into the image automatically using your Brand Kit settings.

  • Powered by GPT Image 2 (OpenAI)
  • Brand logo composited at configured position
  • Brand colours applied to image palette where possible
  • Pro plan: 50 images/month · Max plan: 200 images/month
  • Completely optional. Skip if you source your own visuals
Image Generation (Optional)

Visual Prompt

"Clean scientific illustration of a bioreactor scale-up process, dark blue laboratory aesthetic, minimal line art style, with room for logo in top-right corner"
Logo composited →
🔬

Generating with GPT Image 2…

Using 1 of 50 image generations this month (Pro plan)

Step 08

Copy, review, and publish

Your grounded content is ready. Every claim is tagged, the grounding score is visible, and your source document is long gone from our servers. Review any unverified claims with your team, copy the output, and publish.

  • One-click copy of full text content
  • Grounding tags visible alongside each claim
  • Unverified claims clearly marked for review before publishing
  • Zeplyn adds a factual disclaimer reminder by default
  • Content history saved to your account for reference
Note on regulatory content: Zeplyn generates first drafts. All promotional or HCP-facing content should be reviewed by your medical affairs or regulatory team before publication.
Ready to Publish
✓ 94% Grounded

Bioreactor scale-up shouldn't mean accuracy trade-offs.

Our latest HCP assay validation confirms 99.2% sensitivity across the full manufacturing range, from 10L to 200L, with ICH Q6B thresholds met at every stage. 🔬

✅ 99.2% sensitivity: Grounded (Table 2, p.4) ✅ 10L–200L range: Grounded (Methods, p.6) ⚠️ "shouldn't mean trade-offs": Unverified (opinion framing)

Source: deNOVO_HCP_whitepaper.pdf · Generated 15 Jun 2026 · File deleted ✓

3

Claude AI calls per generation

11

Content types supported

<60s

Average generation time

Try It

Upload your first source document today

See the full pipeline in action. From PDF to published post: grounded, verified, and ready in under 60 seconds.