Overview
The client wanted a SaaS product that helps healthcare businesses produce marketing content: full blog articles, short marketing copy and data on how the content performs, all on one platform. Machine learning supplies topic ideas and checks, and the user stays in charge of what is published.
Solution
In eight weeks we built the platform from scratch. The AI/ML SaaS product is designed for the healthcare sector and supports the creation of material for blogs and digital marketing, from topic research to the finished draft.
Technology Stack
Features
- Automated content generation
- Topic recommendations
- SEO optimization
- Plagiarism detection
- Tone analysis
- Reading level evaluation
- Content scraping and cleaning

Outcome
The platform uses AI to draft content for each user's needs and adds tone analysis, reading level evaluation, plagiarism detection and SEO optimization. Users enter a query and receive topic recommendations drawn from medical databases and search data.
project timeline
built-in checks: SEO, plagiarism, tone of voice, reading level
specialists on the team
Results
- Content generation for different digital marketing formats
- Built-in SEO optimization for groups of related content
- Reading level and tone analysis, so content fits the target audience
- Plagiarism checks before a piece goes out
- Performance data collected for the client's analysis and planning of its marketing
How We Build Content Pipelines for Regulated Topics
Health content needs more checks than ordinary marketing copy, so we build the generation as a pipeline of separate steps. Each step can be tested and replaced on its own:
- Research. Sources are collected and cleaned first: medical databases, guidelines and the client's own approved material. Each source keeps its URL, publisher and date, and is embedded for search in PostgreSQL with pgvector or a vector database.
- Outline. The model proposes a structure for the topic and the target audience. An editor can change it before any text is written.
- Drafting. Each section is written from retrieved passages only, with a citation for every medical claim.
- Checks. Separate steps score SEO coverage, reading level, tone and similarity to existing text on the web. A claim check compares each cited sentence with its source.
- Editorial review. A person approves the draft. Medical statements without a supporting source are marked for the editor.
Today we orchestrate these steps with LangGraph and trace each run in LangSmith, so a weak article can be traced to the step that caused it. Models are chosen per step: a large model for drafting, smaller ones for scoring. The product layer is a TypeScript stack with Next.js and Node.js.
Medical Accuracy and Sources
The main risk in AI health content is a claim that sounds right and is not supported. Adding citations does not solve it by itself, because models often cite sources that do not say what the text says.
Published evaluation shows how common this is. In a Nature Communications study of how language models cite medical references, between 50% and 90% of responses were not fully supported, and sometimes contradicted, by the sources they cited. Even a model with web search enabled left nearly half of its responses not fully supported.
For this reason the claim check in our pipelines is a separate step with its own test set. It compares every cited sentence with the passage it cites and blocks publication when the passage does not support it. Dates on sources are checked as well, because medical guidance changes. The final call stays with a human editor, and the system makes that review faster by showing the exact passage next to each claim.
Related reading: our generative AI development services and AI for healthcare pages.