AI Pipeline for Scientific Content

Chapter 1 – Introduction and context

This project was born from a moment of personal observation and frustration. During a mountain vacation, immersed in a landscape that invited calm and reflection, I found myself scrolling through news on my phone. The experience was disappointing: a monotonous sequence of crime news, gossip and political debates devoid of real informative content. Nothing that stimulated curiosity or conveyed a sense of discovery.

From this realization grew the idea of creating a news stream "as I like them": scientific, curious, positive content, written clearly and accessibly, capable of intriguing even young audiences. My goal was not just to inform, but to do so ethically, transparently and stimulatingly, leveraging the potential of Artificial Intelligence to automate what could be automated, without giving up human control where necessary.

The project targets a broad audience: from science enthusiasts to casual readers, to those seeking alternative information that is lighter but not superficial. It is, in essence, my editorial and technological experiment that unites personal passion and the search for innovative solutions.

Chapter 2 – Sources and licenses

Before delving into specific source selection, it's important to acknowledge the fundamental role of Creative Commons in this project. The non-profit organization founded in 2001 has revolutionized the concept of knowledge sharing, creating a system of standardized licenses that allow authors to specify exactly how their content can be reused. Without this legal framework, a project like mine would not have been possible: Creative Commons has made the world of open licenses accessible and comprehensible, democratizing access to scientific and cultural information.

Source selection was one of the most complex and formative challenges of the entire project. Not coming from the editorial sector, I had to thoroughly study the landscape of open licenses, understanding the differences between Creative Commons BY, CC0 and Public Domain, and learning to correctly interpret reuse clauses.

My goal was clear: use exclusively content from sources that guaranteed legal and transparent reuse. For this reason, source selection remained a procedural and manual phase: I verified each source individually, checking not only the declared license, but also the consistency between what was reported and the actual conditions.

Manual preselection allowed me to eliminate at the root the problem of false positives encountered in initial tests with automatic selection, where AI tended to erroneously classify sources as "reusable" when they were not. This approach also highlighted the importance of maintaining a balance between automation and human control in the most delicate phases of the process.

Verified and used sources

Below is the complete table of sources I manually selected and verified for the project, with related links to licenses to ensure full transparency:

This curated selection represents a balance between scientific quality, thematic variety and legal reliability. Each source is monitored periodically to verify that license conditions remain unchanged and that content quality remains consistent with project standards.

Transparent Attribution & Multi-Author Management

A core tenet of ethical scientific communication is the rigorous recognition of the original contributing researchers. Academic papers are inherently collective endeavors, often authored by dozens of researchers across global universities and research institutes.

To fully honor CC BY 4.0 requirements, the ingestion pipeline features a dedicated author extraction and parsing module: it inspects structured metadata (such as dc:creator elements and item.Authors collections in RSS feeds), sanitizes spurious characters, and attaches full co-author credits alongside direct backlinks to primary sources. Every article published on 3CN serves as a transparent gateway to primary science, upholding moral attribution rights and intellectual integrity.

Adopted License: This work is distributed under the Creative Commons Attribution – NonCommercial – ShareAlike (CC BY-NC-SA 4.0) license, reflecting my commitment to responsible sharing and collaborative creativity.

Chapter 3 – GDPR and Impact on Design

The General Data Protection Regulation (GDPR), which came into force on May 25, 2018, has profoundly transformed the European digital landscape, imposing new responsibilities and strict constraints in personal data management. This regulation, created to strengthen the protection of European citizens' fundamental rights, has also had a significant impact on my project, steering me towards innovative solutions and alternatives to traditional data collection mechanisms.

Obligations and Constraints for Websites

GDPR establishes fundamental principles that every website owner must comply with:

Failure to comply with these provisions can result in administrative sanctions of up to 4% of annual worldwide turnover or €20 million, in addition to possible compensation claims from data subjects. For regulatory insights, please refer to the official website of the Data Protection Authority.

Project Impact: Constraints and Opportunities

Rigorous application of GDPR initially represented a significant obstacle for some functionalities I had planned. In particular, two aspects of the project were directly influenced:

Newsletter and email collection: Implementing a newsletter system would have required managing personal data (email, preferences, navigation metadata) with all related compliance requirements: granular consent, detailed information notices, opt-out systems, secure data storage and management of deletion requests. The bureaucratic and technological complexity, combined with compliance costs, made this option unsustainable for an individual project.

Contact form: Even a simple contact form would have involved collecting and processing personal data, requiring implementation of adequate security measures, specific information notices and management of data subject rights. This aspect, apparently marginal, proved to be an additional element of complexity to avoid.

The Solution: Social Media and Alternative Engagement

Rather than seeing GDPR as a limitation, I interpreted these constraints as an opportunity to explore innovative approaches to content distribution and audience interaction. The adopted solution is based on two main pillars:

LinkedIn Integration

I chose LinkedIn as the main platform for automatic content distribution. This choice presents several strategic advantages: LinkedIn autonomously manages all GDPR-related aspects for its users, offers a professional audience already interested in scientific and technological content, and allows natural engagement through comments, shares and reactions. Each article published on the site is automatically shared on LinkedIn with an optimized presentation post, creating a direct bridge between static content and dynamic community.

Dedicated Page for Contacts and Comments

I developed a dedicated page LinkedIn Page (3CN) that functions as an interaction hub, without directly collecting personal data. This page directs users to appropriate social channels for feedback, questions or collaborations, maintaining editorial control on the main site while delegating communication management to social platforms that already have the necessary GDPR-compliant infrastructure.

Blazor Server: Privacy-First Architecture by Design

The search for maximum GDPR compliance has also profoundly influenced the architectural choices of the project, guiding the decision towards Blazor Server as the development framework. This technology stands out for an intrinsically privacy-friendly approach, eliminating at the root the need to manage traditional technical cookies.

Unlike classic web applications, Blazor Server manages the entire application logic server-side, maintaining user interface state through a persistent SignalR WebSocket connection. This paradigm completely eliminates dependence on cookie-based session mechanisms, making the application naturally compliant with the data minimization principles provided by GDPR.

Technical security note: Since Blazor Server does not implement the traditional HTTP request/response paradigm but uses exclusively WebSocket for client-server communication, it is naturally immune to Cross-Site Request Forgery (CSRF) attacks. This characteristic completely eliminates the need to implement UseAntiforgery mechanisms and related validation tokens, further simplifying the architecture and reducing the application's security footprint.

This technological choice has therefore produced a double benefit: on one hand it guaranteed native GDPR compliance without functional compromises, on the other it generated a cleaner, more secure and performant architecture compared to traditional alternatives based on cookies and HTTP sessions.

Passwordless Administrative Security (MagicLink & OTP)

The "zero unnecessary data" philosophy extends directly to editorial administration. To access sensitive review and scheduling dashboards, I eliminated traditional static passwords — prone to data breaches and credential stuffing — in favor of a Passwordless authentication model based on single-use MagicLinks and OTPs.

The system generates high-entropy cryptographic tokens dispatched via secure channels, active for a single session and immediately invalidated upon first use. To neutralize response timing attacks (timing attacks), token validation utilizes constant-time comparison algorithms, enforced with strict rate limiting and secure disaster recovery mechanisms.

Chapter 4 – EU AI Act: Transparency, Governance & Regulatory Compliance

With the formal adoption of Regulation (EU) 2024/1689 (widely known as the European AI Act), the European Union established the world's first comprehensive legal framework for trustworthy, transparent, and human-centric Artificial Intelligence. The regulation enforces a proportionate Risk-Based Approach, categorizing AI systems based on their potential impact on fundamental rights and safety.

3CN Classification under the European AI Act

The 3CN platform operates as a Limited Risk AI System / General Purpose AI (GPAI) for scientific synthesis and outreach. For this category of applications, EU legislation does not impose prohibitive barriers, but mandates strict, enforceable obligations regarding Radical Transparency, Human Oversight, and Copyright Protection (Articles 50, 52, and 53).

The Five Pillars of AI Act Compliance in 3CN

The matrix below outlines the direct alignment between European legislative mandates and our platform's concrete architectural solutions:

Chapter 5 – Human-in-the-Loop (HITL) & Academic Foundations

In the contemporary discourse surrounding Artificial Intelligence adoption, debate is often polarized between total black-box automation ("end-to-end") and purely manual workflows. The 3CN project demonstrates that highest value is unlocked by adopting the Human-in-the-Loop (HITL) paradigm: a systems architecture where algorithms and humans cooperate in a continuous closed-loop feedback cycle, marrying machine computational velocity with human ethical and qualitative discernment.

Why Is It Termed a "Loop" (Closed Cycle)?

Unlike conventional linear workflows (Input → Unsupervised AI Processing → Public Release) where generative hallucinations and misinterpretations propagate directly to readers, a Closed Loop architecture positions the human editor as an active formative node within the process:

Academic Foundations & The arXiv Connection

The theoretical robustness of this methodology is rooted in international scientific literature. One of the seminal reference surveys in this domain is:

It is particularly noteworthy that arXiv (hosted by Cornell University) is both the origin of this foundational research and one of the open-access scientific repositories natively integrated into 3CN under Creative Commons CC BY / CC0 terms. This synergy highlights the circular coherence of our platform: leveraging open science to understand AI, and employing AI to disseminate open science.

Chapter 6 – Operational Workflow & Processing Pipeline

The pipeline represents the operational heart of my project: a continuous and integrated flow that, starting from information gathering, leads step by step to content publication and distribution. In this architecture, AI intervenes at moments when it can best express its analysis, generation and review capabilities, while the most critical phases from legal and organizational perspectives remain under my direct control.

The flow opens with reading and acquiring feeds from sources I have preselected. This phase is intentionally procedural: I have manually verified all sources to guarantee license compatibility and material quality. The system periodically queries RSS feeds or official APIs, acquiring title, abstract, author, publication date, URL and declared license, which you will find at the bottom of each published article.

It is essential to always verify that the generated article is consistent with the reference scientific publication. Generally, original publications are very articulated and detailed studies, which offer valuable insights for subsequent in-depth analysis and represent a source of high-quality content for further editorial developments.

This is followed by license verification and ethical compliance, also procedural. Each acquired element is compared with the list of licenses I have defined as permitted (CC BY, CC0, Public Domain). I do not scrape the complete content: I only use data present in the feed, respecting the author's will and usage conditions.

At this point AI comes into play with thematic filtering and deduplication. The analysis is not limited to keyword checking, but uses semantic understanding models to evaluate relevance to the objectives I have set. Articles about crime news, politics or gossip are excluded, and any duplicates already present in the database are removed.

After passing the filter, AI proceeds to preliminary classification and rating assignment. Each article is evaluated on four dimensions I have defined: positivity, scientificity, novelty and expository clarity. Each criterion is assigned a score from 1 to 10, and the arithmetic mean determines the overall rating.

The WRITER phase is entirely entrusted to AI. Here the content is rewritten following the editorial rules I have established: clear and accessible language, positive and curious tone, absence of personal opinions, rigorous respect for licenses and text originality. The objective I have set is to transform a technical abstract into a popular science article that maintains information accuracy but makes it accessible to a broad audience.

When the quality score does not reach the minimum threshold I have established, the REVIEWER phase comes into action. Here too AI is the protagonist, but with an iterative approach I have designed: the text is reworked multiple times, up to a maximum of ten cycles, trying to correct weaknesses and improve evaluation parameters.

The Neural Voice Synthesis (TTS) phase converts the approved article into a high-definition audio track for mobile listening (with automatic client-side fallback to the browser speech engine). For the full technical breakdown on neural models, the Voice Gallery, and dual-engine architecture, see Chapter 8: Neural Voice Synthesis.

The Neural Vector Indexing (ONNX) phase automatically computes 384-dimensional dense embeddings for every new article, storing them in SQL Server database to enable instant conceptual searching. For full architectural breakdown and performance benchmarks, see Chapter 11: Neural Semantic Search.

The cycle closes with publication on the site in dual language (Italian and English), dynamic generation of aggregated RSS feeds, and automated sharing on LinkedIn, expanding distribution reach and engaging an interested professional audience.

Chapter 7 – The orchestrator prompt: architecture and operational logic

The beating heart of the entire system I developed resides in a single complex prompt, the result of numerous iterations and refinements. This component, apparently simple in its textual form, actually represents the most critical and innovative element of my project: a true "orchestra conductor" capable of coordinating distinct roles, managing complex decisional flows and guaranteeing structured and coherent outputs.

The design of this prompt was an evolutionary process in my journey. The initial approach I had adopted involved separate calls for each phase, but this system proved inefficient on multiple aspects, requiring 60-70% longer times and complex procedural management of various intermediate outputs.

The tripartite structure I developed

STEP 1 – WRITER defines the editorial identity of my system. Here I established the primary objectives: transform scientific and technical content into texts accessible to young audiences, maintaining rigor and accuracy but adopting a fresh and engaging tone. The editorial rules I defined are particularly detailed because they represent the project's DNA: absolute originality to avoid copyright issues, neutrality to guarantee credibility, expository clarity to reach the target.

STEP 2 – REVIEWER introduces a self-evaluation and iterative improvement mechanism I designed. The system assigns scores from 1 to 10 on four dimensions I chose – positivity, scientificity, novelty and expository clarity – and calculates an average that determines whether content is acceptable or needs reworking.

This section includes safety filters I implemented to automatically exclude content not consistent with my editorial line: non-scientific articles, promotional material, sensitive topics like serious diseases. The improvement cycle can repeat up to ten times, a limit I established to avoid infinite loops but sufficient to allow significant optimizations.

Structured output management

STEP 3 – OUTPUT represents the most significant innovation I introduced from a procedural integration perspective. Instead of returning free text, the system produces a structured JSON object with predefined fields, each optimized for a specific function within the pipeline I created.

The JSON structure I designed includes elements in dual language (Italian and English), URL slugs optimized for SEO, hashtags calibrated for social media and a "BonusExpert" section with technical insights for more advanced readers. Each field has specific compilation instructions I defined: length, tone, style, technical characteristics.

A fundamental aspect is the rating threshold calibration I had to optimize through testing. In my initial tests, thresholds too high (above 8.0) produced an excessive number of discards, while thresholds too low (below 6.0) let through content of insufficient quality. The threshold I currently use, dynamically configurable, represents an optimal compromise between quality and productivity I achieved through experience.

The orchestrator prompt I developed therefore represents much more than a simple textual instruction: it is an integrated quality control system, an automatic editor and a specialized translator, all condensed into a single conversational interface. Its design required months of refinements, tests and corrections, but the final result is a powerful, reliable and scalable tool.

Chapter 8 – Neural Audio TTS & Voice Gallery Integration

One of the most rewarding and impactful extensions of the project has been its multimodal evolution: upgrading the platform from a text-only news aggregator into an immersive augmented listening experience. In today's dynamic information landscape, on-the-go consumption — while driving, exercising, commuting, or taking screen-free breaks — has become paramount. However, user satisfaction hinges on natural vocal timbre, expressive cadence, and precise pronunciation of specialized scientific vocabulary.

Bridging the Gap: From Robotic Browser Voices to Deep Neural Models

Traditional client-side synthesis via the browser Web Speech API suffers from notable shortcomings: mechanical and monotonous delivery, heavy reliance on local voices installed on the user's operating system (leading to stark discrepancies across iOS, Android, and Windows), and frequent mispronunciations of compound words or scientific terminology.

Azure AI Speech Ecosystem & Microsoft Voice Gallery

The selection and benchmark of voice profiles centered on the cloud capabilities of Azure AI Speech and the interactive Azure AI Voice Gallery , showcasing hundreds of studio-quality 48 kHz neural voices across global languages.

Following comprehensive phonetic stress testing against AI-generated content, I calibrated a dedicated, high-clarity neural voice profile for each language, powering both the main summary and the Expert Bonus section:

* Note on progressive migration: Because high-definition (HD) neural audio generation requires dedicated server-side rendering time and compute resources, the migration process is being rolled out incrementally: it is currently active and prioritized for Italian content, and will subsequently be extended to the global English edition.

Dual-Engine Architecture: High-Definition & Resilient Fallback

To combine studio audio quality with operational efficiency (zero client-side download overhead and controlled cloud costs), I engineered a Dual-Engine pipeline strategy:

1. Neural HD Layer (Server-Side Batch Pre-Rendering): During the editorial review workflow, the admin can trigger high-definition neural synthesis. The backend interfaces with Azure AI Speech, downloads the optimized MP3 stream, and stores it in static CDN storage. End users enjoy zero-latency, instant playback with a distinct ✨ HD badge.
2. Client-Side Fallback Layer (Web Speech API): For articles without pre-rendered audio or in offline contexts, the Blazor player automatically falls back to the native browser speech engine with a 💻 TTS indicator, ensuring 100% service availability under any network condition.

Chapter 9 – Software Architecture & Technology Stack

I evaluated various technological architectures, initially orienting myself towards Blazor WebAssembly integrated with Web API. However, this approach revealed rigidity incompatible with Agile methodology: refactoring times necessary to implement direction changes often exceeded those required for business logic development. Consequently, I opted for migration to Blazor Server, which offers greater flexibility and more sustainable development times.

The strategic choice of Blazor Server 8

My decision to adopt Blazor Server 8 as the development framework was profoundly influenced by the methodological approach I wanted to follow. Knowing I had to face an experimental project with many unknowns, I opted for Agile management that would allow me to "do and redo" various parts of the system iteratively and reactively.

Blazor Server 8 allows me to build interactive web applications using C# and the entire .NET ecosystem, eliminating the need to resort to multiple stacks and different languages. This choice allowed me to maintain a coherent development environment, drastically reducing cognitive complexity and accelerating implementation times for changes.

Real-time and interactivity

SignalR integration enables real-time updates that would have been much more complex to implement with an API-based architecture: batch progress status, arrival of new articles or modification of scheduling become immediately visible without reloading the page. This functionality, crucial for monitoring an automated system, was implemented in a few lines of code.

Using Blazor Server has significantly simplified cache management for today's articles. Since this content remains unchanged during the day, it's inefficient to continuously query the database. A scheduled event at midnight expires caches and loads new articles; simultaneously, a trigger connected to this event uses APIs to automatically publish a presentation post on LinkedIn.

From a visual experience perspective, the interface I designed is responsive and optimized to work on desktop, tablet and smartphone. The dynamic theme system I implemented allows real-time switching from light to dark mode, adapting to user preferences. Multilingual support I implemented leveraging .NET's native functionality, allowing me to offer content in Italian and English without duplicating logic or interfaces.

Editorial Review Board Dashboard

The operational core for Human-in-the-Loop oversight is the private Editorial Review Console. This advanced interface empowers the administrator to inspect every AI-processed article in real-time before public release:

• Multi-Metric Inspection: Detailed breakdown of quality scores (Positivity, Scientificity, Novelty, Clarity) and verification of extracted original author credits.
• Rapid Decision Workflow: Capability to approve articles for scheduled publishing, discard them permanently, or trigger guided AI re-elaboration with custom prompt instructions.
• Direct Editorial Editing: Fine-grained editing of bilingual titles, body summaries and bonus insights to polish phrasing and tone.
• One-Click Neural Audio Generation: On-demand dispatching of high-definition TTS synthesis with instant MP3 file pre-rendering.

Quartz.NET Batch Scheduler & Live Terminal

Asynchronous background job orchestration is managed via a dedicated Batch Control Console, powered by Quartz.NET with persistent SQL Server storage. Through this operational console, administrators can inspect scheduled nightly batch runs, adjust cron schedules live, launch extraordinary batch cycles with a single click, and observe streaming runtime logs directly on screen.

Ultra-Lightweight Custom ORM (DbCore Pattern)

To maximize read/write throughput and prevent external ORM bloat, I engineered a custom ORM framework adhering to the DbCore pattern. Utilizing declarative annotations ([DbEntity], [DbKey], [DbColumn]), the architecture ensures strongly-typed mapping, optimized SQL queries, built-in injection safety, and minimal memory footprint across domain entities and audit trails.

Dynamic AI Model Management & Automatic Fallback

A standout architectural element is the integration with OpenRouter featuring a dynamic fallback strategy: should the primary model (such as mistral-small-3.2-24b) become saturated or hit rate limits, the pipeline automatically fails over across high-performance alternatives (gpt-oss-120b, llama-3.3-70b, gemini-2.0-flash), guaranteeing 100% operational continuity.

Observability and Cloud-Ready Scalability

The architecture embeds enterprise-grade logging (NLog) tracking every operation: LLM API latency, RSS HTTP response codes, voice synthesis telemetry, and an immutable audit trail of all editorial approvals. The platform is architected for containerization and horizontal scaling, fully leveraging .NET 10 performance and SQL Server reliability.

Chapter 10 – Accessibility & Digital Inclusion (WCAG 2.1 AA)

The 3CN platform is designed following the principles of universal accessibility (W3C WCAG 2.1 Level AA and European Accessibility Act), ensuring scientific knowledge is fully accessible to everyone, including blind and visually impaired users, people with dyslexia, or motor disabilities.

Chapter 11 – Neural Semantic Search & ONNX Vector Embeddings

Conventional search engines for blogs and news portals rely historically on exact lexical matching (traditional keyword lookup or BM25 algorithms). If a user searches for "non-polluting vehicles", a traditional engine returns zero articles if the titles contain only the terms "electric cars" or "hydrogen mobility". To bridge this conceptual divide, the 3CN platform integrates an in-process neural semantic search engine based on dense vector embeddings and high-performance inference with Microsoft ONNX Runtime.

1. The Neural Model: all-MiniLM-L6-v2

The algorithmic core of the search is the all-MiniLM-L6-v2 model, developed within the open-source Sentence-Transformers framework based on Microsoft's distilled Transformer architecture. With 22.7 million parameters (approx. 0.023B), a 6-layer self-attention structure, and a hidden dimension of 384 units, the model is specifically trained to map sentences and documents into a continuous 384-dimensional vector space.

2. INT8 Dynamic Quantization: Maximum Efficiency at Zero Cost

The original 32-bit floating-point (Float32) Transformer model occupied approximately 120 MB of RAM. To enable in-process execution on shared-host servers without degrading application load times, we performed dynamic 8-bit integer quantization (INT8) using ONNX Runtime's official toolkit (onnxruntime.quantization.quantize_dynamic).

3. Vector Storage & Ultra-Fast Cosine Similarity

The 384-dimensional vectors computed for each article are L2-normalized ($\|v\|_2 = 1$) and stored directly in the relational database as a compact binary stream (VARBINARY of exactly 1,536 bytes per article). The entire historical news archive takes up just 3.5 MB of total vector storage (less than 0.35% of database quota).

Thanks to prior L2 normalization, Cosine Similarity simplifies mathematically to a simple Dot Product (scalar product), enabling the engine to compare a query against thousands of articles in under 2 milliseconds on a single CPU core.

4. On-Demand Lifecycle & Sliding Lease Memory Manager

To maximize system memory efficiency on production environments, the semantic engine uses an on-demand sliding expiration lease pattern:

5. How to Test Semantic Search on the Portal

You can experience neural conceptual search live directly in the search bar of the /news page:

6. Search Comparison: Neural AI vs Exact Keywords

The following table illustrates a side-by-side comparison (2 neural AI conceptual matches and 2 exact keyword matches):

In the era of Artificial Intelligence, online discoverability is undergoing the most radical transformation since the inception of the World Wide Web. Traditionally, Search Engine Optimization (SEO) revolved around competing for the top ten blue links on Google by optimizing keywords and acquiring backlinks. Today, this landscape has shifted toward Generative Engine Optimization (GEO): millions of users no longer search by typing isolated keywords, but interact through natural dialogue with generative answer engines such as ChatGPT Search, Perplexity, Claude, and Google Gemini.

For a scientific dissemination platform, the goal is no longer just "getting found", but being architected to be understood, synthesized, and cited as an accredited primary source by AI models, all while achieving instantaneous real-time indexing. To achieve this, MantoX Blog 2 engineered a four-tier architecture combining modern semantic web standards, real-time push indexing protocols, and native data structures tailored for Large Language Models.

1. The Real-Time Revolution of IndexNow: Search Engines' Best-Kept Secret

For decades, search indexing operated on a Pull model: a website publishes a new story, appends it to an XML sitemap (sitemap.xml), and waits passively. Search engine bots (web crawlers) must periodically revisit the sitemap, compare timestamps, and decide when to crawl. This round-trip often takes days or even weeks, introducing an unacceptable delay for fast-moving science and breaking breakthroughs.

MantoX Blog 2 implements the open IndexNow protocol, spearheaded by Microsoft Bing and supported by a growing network of global engines (including Yandex, Seznam.cz, and Naver). While still relatively unknown to the general public, IndexNow is recognized by web architects as one of the most substantial search infrastructure breakthroughs in recent years.

2. GEO Standards `llms.txt` & `llms-full.txt`: Speaking the Native Language of LLMs

While traditional search engines parse web pages through HTML tags and CSS layouts, Large Language Models (such as ChatGPT, Claude, and Perplexity) require dense, semantic, noise-free text that is easy to ingest. Following the open llmstxt.org specification, MantoX Blog 2 serves two dynamic, lightweight endpoints:

Thanks to this architecture, when users ask a generative AI about a recent scientific discovery or physics paper covered on MantoX Blog 2, the AI can directly retrieve the context, cite our platform as a verified source, and provide accurate academic attribution.

3. Structured Semantic Graph: High-Precision Schema.org `@graph`

To communicate unambiguously with Google News algorithms and semantic parsers, every article dynamically injects a unified 5-dimensional JSON-LD (Schema.org) graph in the <head>:

• NewsArticle: Formalizes article metadata, bilingual headlines, authorship, ISO 8601 publication and modification timestamps, and certified publisher data.
• AudioObject: Links the neural HD podcast MP3 track with technical audio bitrate, encoding, and estimated playback duration.
• SpeakableSpecification: Instructs voice assistants (Google Assistant, Alexa, Siri) and screen readers which CSS regions (title, summary, bonus insights) should be read aloud during audio playback.
• BreadcrumbList: Defines semantic navigation hierarchy (Home › Category › Article), allowing search engines to display rich breadcrumbs in search results.
• FAQPage & Key Takeaways: Encodes interactive "Smart Q&A" pills (Explain it simply, Pros & Cons, What's ahead) as questions and answers, enabling Google SERP expandable rich snippets.

4. Dynamic OpenGraph Social Preview Engine (`/og-image/{slug}`)

Social media sharing (WhatsApp, Telegram, LinkedIn, 𝕏, Facebook) is a primary distribution channel for digital publishing. Rather than relying on generic fallback images, MantoX Blog 2 features a dedicated OgImageController that generates customized, high-definition OpenGraph preview cards on-the-fly for every article.

Each card features the article title, category badge, estimated reading time, neural podcast indicator, and the portal's signature visual gradient, ensuring high visual clarity across all devices and social feeds.

5. Editorial Transparency & EEAT Trust Signals

Google's EEAT (Experience, Expertise, Authoritativeness, Trustworthiness) quality rater guidelines and modern AI models reward radical transparency in digital publishing. On our platform, every article features a prominent Trust Box displaying:

Looking back at the entire journey, this project has proven to be a voyage as challenging as it was formative, where each phase represented a technical, methodological and personal challenge. My initial idea of entrusting Artificial Intelligence with the entire pipeline, from source selection to publication, clashed with the reality of a domain – the editorial one – that requires particular attention to content provenance and legal compliance.

Fully automatic source management showed evident limitations in my tests: too many false positives in license classification, with sources erroneously labeled as reusable, and an unsustainable expenditure of time and resources. This led me to a crucial decision: adopt a hybrid model, where source preselection and editorial planning remain under my procedural control, while AI intervenes in steps where it can express maximum added value.

AI Brainstorming

Collaborative brainstorming with AI proved to be the most surprising aspect of the project: the ability to share ideas and receive increasingly relevant and stimulating suggestions transformed the methodological approach, favoring design over immediate development. Preventive analysis of complex scenarios enabled a significant reduction in overall project time, avoiding the improvised approach typical of "vibe coding" that often leads to dead ends or unsatisfactory solutions. AI Copilot proved to be a valuable assistant, while the mobile application became an almost indispensable tool for sketching ideas to develop later or conducting in-depth research in real-time.

AI Code Production

An aspect that deserves specific reflection in my experience is the use of AI for code generation. The initial approach I had attempted, based on "block" methods like vibe code, proved inefficient: too much redundant code, technical and functional anomalies, and revision time often exceeding that necessary to write code manually.

I identified a recurring flaw of "amnesias" due to token limits in responses and, especially, the fact that models, having limited memory, tend to deliberately discard content deemed less relevant. When subsequently requested, this content is often reinvented or presented through superficial summaries, generating frequent errors and inconsistencies. This limitation makes constant and careful supervision of all AI-generated outputs indispensable.

The personal dimension of my project

On a personal level, I developed the project mainly during nighttime hours, after putting my daughter to bed. My wife had established an absolute ban on touching the PC after dinner, so I started intensively using the Copilot mobile app to design and develop ideas. This "suspicious" behavior with the smartphone made her suspicious to the point of checking on me, convinced I had a digital lover! The irony is that, in a certain sense, she was right: my relationship with AI had become quite intense. Despite these amusing domestic obstacles, seeing a complex and functional system take shape day by day, born from a casual idea in the mountains, proved extremely satisfying and made every nighttime sacrifice definitely worth it.

The software development part also presented its challenges in my experience. Adopting Blazor Server and .NET 10 as framework, initially difficult due to my limited familiarity and the fact that this version had broken compatibility with some previous patterns (creating considerable confusion), transformed over time into a strength. Consistency with the .NET ecosystem and the ability to naturally integrate complex functionality proved to be valuable characteristics that more than compensated for initial difficulties.

Towards a replicable model: The New Frontiers of MantoX Blog 2

In conclusion, this experience has reinforced a conviction: the hybrid model, where AI is used in a targeted and non-totalizing way, can be the key to integrating artificial intelligence into business and creative processes. It's not about "doing everything with AI," but using it where it really serves, preserving human control at critical points.