Automating content creation with generative AI

Consider a company publishing a daily industry briefing: three important stories, each with a short description, accompanied by one shared illustration.

The system must independently find events within a defined period, identify significant news, remove duplicates, check sources, prepare text, create an image and publish the finished content.

A single pipeline connects web research, ranking algorithms, generative AI models, software checks and publishing channel integrations.

For example, an automated scheduled process gathers news each day from RSS, search services, official websites and other permitted sources.

python-script receives headlines, descriptions, links, dates and available publication text. Results are stored in PostgreSQL for further processing.

The system then identifies publications describing the same event. Materials from multiple sources are grouped together, so repeated news does not occupy several places in the briefing.

Intelligent selection follows. The algorithm assesses relevance, novelty, importance to the audience, source quality and compliance with the chosen criteria.

The three most suitable events are selected from the verified candidates.

Using structured information, text-LLM-model creates the final text: a headline, short event descriptions and the required source links.

In parallel, or after the text brief is prepared, image-LLM-model creates an original thematic illustration matching the content and design requirements.

Generated content goes through automated validation. The system checks claims against sources, text length, required elements, duplicates, style and technical image parameters.

If an error is found, the material returns for automatic correction. If the primary algorithm cannot complete the publication, the system automatically switches to a fallback publishing algorithm. The finished content is published after validation.

Once all checks pass, the pipeline sends the content through an API to Telegram, a website or another target system.

PostgreSQL records the sources used, content versions, check results, publication time and final execution status.

The entire cycle from news research to a finished publication can run automatically, without manual copying, editing or content assembly.

1. Automated start and source material collection

The process starts on a schedule, through an API, a webhook or another event.

For example, the system can produce daily news, create a product description when an item is added, or prepare a sales proposal when a CRM deal changes.

python-script receives source information through permitted integrations: Web Search, RSS, APIs, PostgreSQL, CRM, ERP, file storage and other sources.

Data are converted to a common structure and stored for further processing.

The pipeline can record the collection window, content category, language, audience, source requirements and other task parameters.

2. Filtering, deduplication and intelligent selection

Not every collected item is suitable for content generation.

Before producing content, the system cleans data, removes duplicates, groups related materials and checks their relevance to the chosen topic.

Complex selection can use text-LLM-model, semantic matching, embeddings and ranking algorithms.

For example, several news websites may describe the same event differently. The system identifies semantic similarity, groups the materials and prevents duplication in the final briefing.

Candidates are then assessed against the defined criteria. Depending on the task, these can include relevance, novelty, importance, verifiability, commercial relevance or a particular customer's requirements.

The main principle is that generation starts with prepared, selected information rather than arbitrary production of plausible text.

3. Generating text and structured materials

Selection produces a brief for text-LLM-model.

It specifies the source facts, required structure, language, style, length limits and mandatory elements.

Depending on the task, the model can create:

• news posts and articles;
• product and service descriptions;
• website and social media copy;
• sales proposals;
• personalised messages;
• document summaries;
• analytical briefings;
• materials for reports and presentations.

When a defined structure is required, the model returns JSON or another agreed format.

Then python-script uses these data to create HTML, Word, PDF, email or an API publication.

Documents with fixed details, calculations and fields are best produced by combining an LLM with software templates. Numeric calculations and required details should not be left to unrestricted text generation.

4. Automated image generation

If the business process requires visuals, it uses image-LLM-model.

The system prepares an illustration brief from the topic, verified facts and design requirements.

The image may be intended for a Telegram post, article, product page, advertisement or internal document.

The pipeline specifies the format, dimensions, composition, visual elements and constraints.

After generation, the image is checked against technical requirements and the chosen topic.

News and analytical materials should distinguish an original conceptual illustration from a documentary photograph of an actual event. An image should not falsely present an imagined scene as an authentic news photograph.

5. Automated validation and correction

Generative models can produce persuasive text containing errors, so automated generation must be accompanied by independent checks.

Standard software logic handles some checks:

• text length and format;
• required fields;
• compliance with the defined structure;
• repeated fragments;
• technical image parameters;
• permitted file sizes;
• valid addresses and identifiers;
• absence of previously published duplicates.

Other checks require semantic analysis.

A separate text-LLM-model or a dedicated validation stage can compare claims with the supplied sources, find unsupported wording and identify contradictions.

When a problem is found, the system provides specific feedback and sends the material for correction.

The corrected content is checked again.

Specific feedback guides text regeneration towards correcting the detected errors. A suitable prepared image is available if generation or moderation prevents use of a new illustration. Once required checks pass, the issue is published automatically.

Factual claims are compared with primary sources. Unsupported details are removed or replaced with accurate wording so the corrected text retains the events' meaning and matches the original materials.

6. Publication and delivery

After all mandatory checks pass, the system delivers the result to its target destination.

This can include:

• publishing a Telegram post;
• creating a website article;
• sending a personalised email;
• generating Word or PDF;
• updating a product page;
• creating a CRM record;
• passing data to another system through an API.

Confirmation from the destination system marks the operation complete.

If publication fails because of a temporary API problem, the pipeline preserves its state and can retry safely.

Operation IDs, status checks and logging prevent duplicate publication.

7. Monitoring, history and execution control

Each run can save its history: collected materials, selection results, generation parameters, text and image versions, validation feedback, errors and the final status.

PostgreSQL manages states, metadata and technical history.

This supports process analysis, quality and generation cost monitoring, identification of repeated errors and improvement of the rules.

Generative AI becomes part of a repeatable, manageable software process.

Automated product descriptions

A new product appears in ERP → the system receives its attributes → creates a name, description and supporting content → checks facts and format → updates the catalogue.

Personalised sales proposals

CRM supplies customer and deal details → the pipeline selects current services and prices → an LLM prepares proposal text → software creates Word/PDF from a template → the document is checked and delivered along the approved route.

Automated reports and analytical briefings

PostgreSQL and BI supply calculated metrics → an LLM produces a clear explanation → the system adds tables and charts → checks numeric data → creates the final document.

Content for multiple channels

One verified material supplies several formats: a website article, a short Telegram post, a customer email and social media copy.

Each version follows the channel's requirements without manually creating the content again from scratch.

Personalisation for a customer or audience

Using permitted data and segmentation rules, the system adapts content, language, structure and presentation.

Personalisation stays within the available permissions and must not disclose someone else's data.

Automated generative processes reduce manual operations in content preparation and distribution.

One pipeline performs information research, copying, writing, illustration creation, validation and publication.

The system can run scheduled tasks, process events, create several formats and send them to the required channels.

Consistent generation and validation rules maintain a shared style, enforce requirements and record completed operations.

People retain control: they define the task, permitted sources, quality criteria, publishing rules and situations requiring additional approval.

The main value is automation of the whole sequence needed to obtain a finished result, beyond generating an individual text or image.

Depending on the task, the system uses:

Python — orchestration, processing and software checks.

PostgreSQL — data, history, states and results.

Web Search / RSS / API — source information collection.

text-LLM-model — text generation, classification, selection and semantic validation.

image-LLM-model — original illustration creation.

HTML / Word / PDF — content preparation.

Website / Telegram / email / CRM — publication and delivery.

The architecture depends on run frequency, material volume, quality requirements and the complexity of automated actions.

One AI model can write text. Another can create an image.

A more useful business system independently gathers data, selects the relevant information, coordinates several models, checks the result and delivers finished content where it will be used.

Research → Selection → Generation → Validation → Publication

This approach makes generative technologies a complete automation tool.