RomânăOncoExpertAI — Summary with Focus on Technology and Technical Advantages
Documentation for Specialists
Overall Technological Vision
OncoExpertAI is not a simple medical chatbot, but a complex hybrid platform that solves a fundamental problem of AI applied to medicine: how to get accurate, empathetic and error-free answers in a field where wrong information can cost a life. The proposed solution combines a private medical knowledge base, unique in the world, with a multistep processing system and state-of-the-art AI models, all orchestrated through an advanced RAG architecture.
The Fundamental Technical Problem It Solves
Before understanding the solution, it is essential to understand why existing commercial AI models fail in the medical context. The paper identifies three critical technical limitations of them.
The first is limited processing capacity — generic models cannot handle large volumes of medical documents (60–120 pages of clinical records) without losing coherence or omitting essential clinical details. The second is the phenomenon of hallucination — AI models easily generate false information with plausible appearances, which is unacceptable in medicine. The third is the lack of personalization and empathy — answers are too short, impersonal, and lack the full medical context of the patient.
The OncoExpertAI system was specifically designed to eliminate each of these limitations.
Central Architecture: Multistep Hybrid System
The main technical innovation is a hybrid multistep/multimodule architecture , in which each complex medical problem is decomposed into distinct logical steps, each step benefiting from dedicated specialized interventions. This approach is superior to a single prompt sent to a generic model, because each step filters, verifies, and enriches the information before passing it on.
The processing flow comprises six successive stages.
Stage 1 — Data Analysis involves the system processing massive volumes of medical documents, including poor quality scans, extracting and structuring the information while preserving all relevant clinical details.
Stage 2 — Classification and Categorization automatically identifies the patient type and area of medical interest to correctly guide the entire subsequent process.
Stage 3 — Module Selection chooses the specific medical module relevant to the case from the knowledge base, loading the validated context for evaluation.
Stage 4 — Guided Dialogue (Advanced RAG) represents the core of the architecture, with the patient interacting with an AI constrained to respond exclusively based on validated information from the chosen module.
Stage 5 — Personalized Output generates clear and empathetic medical answers, adapted to the patient’s level of understanding, with simplified complex explanations.
Stage 6 — Continuity ensures that the dialogue can continue indefinitely, and the information remains available in written format for further consultation.
Advanced RAG Mechanism — The Anti-Hallucination Solution
The most important technical element of the platform is the implementation of advanced RAG (Retrieval-Augmented Generation) , which represents the fundamental difference from any other existing medical AI system. Through this architecture, the AI model does not “invent” information from its general training, but is technically constrained to respond exclusively from validated medical modules included in the active context of the session. Basically, the AI functions as an expert who can only read and interpret the documents placed on its table, not from its own potentially erroneous memory.
The first technical pillar of this mechanism is semantic matching , which is the system’s internal search engine. When a patient describes a symptom in common or colloquial language, the system does not search through the 230,000+ pages of the knowledge base by simply matching keywords — a classic, rigid and imprecise method. Instead, it uses vector representations of the text (embeddings) to understand the meaning and intent of the question, correlating it with the most relevant module even when the patient’s terminology differs completely from the formal medical one. This capability is essential in an oncology context, where patients often describe complex symptoms in their own words, and the system must correctly and quickly identify the right module without losing important clinical nuances.
The second pillar is the in-context learning mechanism , whereby the AI model effectively becomes a temporary specialist dedicated to the patient’s specific case. Once the relevant modules are selected and loaded into the active session context window, the model no longer operates from its general training knowledge — which may be incomplete, generic, or outdated — but learns on the fly exclusively from the validated medical material provided in that particular context. This dynamic adaptation means that each session is, technically, a consultation with a recalibrated expert specific to that patient’s profile, with direct access to the most relevant and up-to-date protocols for their case.
The combination of the three mechanisms — RAG constraint, semantic matching, and in-context learning — virtually eliminates hallucinations, the number one problem of AI in medicine, while maintaining empathy and naturalness in communication. The result is a system that is simultaneously clinically accurate, up-to-date with the latest international protocols, and human in tone — a combination that no generic AI model or standard consultation can consistently replicate.
The Medical Database — A Unique Technological Asset
The foundation of the entire system is a private medical knowledge base, described as unique in the world in that it was built specifically to be compatible with artificial intelligence processing, not for human reading. This is a critical technical distinction: the information is structured, segmented, and labeled so that AI models can index, retrieve, and use it with maximum precision.
The size and coverage of this database are impressive: over 3,000 specialized modules developed and tested, a total volume of 230,000-250,000 A4 pages (the equivalent of hundreds of updated medical treatises). The modules are organized into 4 main categories;
- specific oncology modules (approximately 600)
- general interest modules on anatomical regions and symptoms,
- disease-specific modules with detailed information about treatments and side effects, and
- general modules by specialty for cases with unclear diagnosis or associated diseases.
Medical coverage is exhaustive, including complete oncology with TNM classification and international guidelines, all major non-oncology specialties (cardiology, neurology, endocrinology, gastroenterology, hematology, pulmonology, nephrology, dermatology, infectious diseases and others), rare diseases with dedicated chapters, plus specialized sets for known and research-based genetic mutations identified from clinical trials, published articles, etc.
Selected AI Models and Reasons for Choice
A deliberate and well-reasoned technical decision was to select three complementary AI models instead of just one: Google Gemini 2.5 Pro, 3.1 Pro , and Anthropic Claude 4.6 Sonnet . This choice is not accidental — the 3 models are used to ensure the optimal balance between three critical dimensions: computational performance and long-term context processing capacity (Gemini 2.5 Pro and 3.1 Pro), respectively, empathy in communication and accuracy in human-nuanced responses (Claude 4.6 Sonnet).
Using three models in tandem allows the system to compensate for each one’s weaknesses and capitalize on their strengths, resulting in consistently high quality over long sessions and an empathetic tone tailored to the real emotional needs of the oncology patient.
Medical Document Processing — Superior Technical Capability
One of the concrete technical advantages of the platform is the ability to process massive volumes of medical documents , including poor quality scans. The system automatically extracts and structures the information, correlates laboratory data with the patient’s history and available imaging information, checks treatment compatibility and identifies potential drug incompatibilities, and generates synthetic reports of up to 20+ pages equivalent to what a senior consultant would produce after several physical consultations.
This capability directly solves one of the biggest frustrations of oncology patients: the practical impossibility of presenting a complex medical file to the treating physician and receiving an integrated and complete analysis within a reasonable time.
Integration of International Medical Sources
Technically, the knowledge base is not built in isolation, but integrates and aligns information from the most important international medical sources and protocols: ESMO, ASCO, NCCN and AJCC protocols for therapeutic guidelines, PubMed with over 40 million scientific articles, MedlinePlus, Europe PMC, and the Annals of Internal Medicine (ACP) for extensive medical literature; and ClinicalTrials.gov, EU Clinical Trials, Cancer Currents Blog (NCI) for active clinical trials.. This integration means that each answer generated is essentially the equivalent of synthesizing thousands of validated sources simultaneously — something impossible to achieve in a traditional 15-minute consultation.
Technical Multilingualism
Another important technical decision is that all modules are written in English — the dominant language of international medical literature — with built-in machine translation for dialogue with patients in any language. This approach ensures that there is no loss of medical accuracy through the translation of sources, while maintaining accessibility for patients in any country.
Integrated Oncology — Complete Medical Solutions, Not Isolated
One of the fundamental limitations of traditional medical systems is the tendency to treat cancer in isolation, ignoring the patient’s concomitant diseases. An oncology patient often also has diabetes, hypertension, renal failure, cardiovascular or liver diseases, and these comorbidities directly influence the choice of oncology protocol, treatment tolerance and overall prognosis. That is why the OncoExpertAI knowledge base includes extensive modules for all major medical specialties — cardiology, endocrinology, nephrology, hepatology, neurology, gastroenterology, pulmonology and others — not as a simple additional coverage, but as a deliberate architectural necessity. The system can thus correlate in real time the oncology diagnosis with the patient’s entire medical profile, identifying interactions between chemotherapy and cardiac medication, contraindications generated by reduced renal function or dose adjustments imposed by pre-existing liver disease. The result is a truly integrated analysis, in which the proposed oncological solution takes into account the patient’s entire clinical reality, not just their tumor — exactly as a top multidisciplinary tumor board would do, but accessible anytime and to anyone.
Technical Security and Privacy
At the infrastructure level, the platform uses SSL encryption for data protection in transit, secure storage of medical documents with strictly controlled access, full compliance with GDPR and international medical legislation, and the possibility of completely anonymous use — an important differentiator compared to any traditional consultation.
From an architectural standpoint, it is essential to emphasize that the OncoExpertAI platform functions exclusively as a secure and intuitive interface for communicating with the patient (frontend). All complex multi-step processing algorithms, advanced RAG mechanisms, the knowledge base, and the AI models that analyze medical data are not hosted directly on this user-facing platform but run on a separate, dedicated, and highly secure backend infrastructure. This decision to decouple the technical components ensures not only a seamless user experience but also the protection of intellectual property and maximum data security by separating the patient’s direct interaction from the massive computing power required for clinical processing.
Human Validation (Human-in-the-Loop) — Absolute Security Guarantee
Although, from a strictly technical point of view, AI models generate answers in a few minutes (even when analyzing hundreds of pages through a complex multistep flow), the delivery time of the final report to the patient is set between 3 and 48 hours for an essential clinical reason. This interval does not represent a system speed limitation nor delayed processing to save computing resources, but the deliberate integration of a critical safety mechanism: the “Human-in-the-Loop” approach . Specifically, to guarantee absolute accuracy and eliminate any residual risk, each analysis and recommendation generated by artificial intelligence is personally reviewed, validated and approved by Dr. Onisim Florin before being delivered. Thus, the speed and synthesis capacity of the technology are doubled by the discernment, experience and responsibility of human medical expertise, ensuring that the patient receives 100% correct and safe information.
Validated Technical Results
The system has undergone a testing period including on an independent platform , with 100% positive feedback. Demonstrated results include generating over 20 pages worth of analyses with detailed accuracy, high consistency of quality over long sessions, consistently maintained empathic tone, and the practical elimination of hallucinations through the RAG architecture.
Technical Competitive Advantage — Final Synthesis
Compared to traditional consultation and generic digital platforms, the technical advantage of OncoExpertAI can be summarized as follows: the system combines the specificity of a human expert (modules validated by Dr. Onisim Florin and international protocols), the scalability and speed of AI (response in 3–48 hours, available 24/7, no waiting lists) and architectural safety (advanced RAG that eliminates hallucinations). No existing solution on the market combines all three dimensions simultaneously, which is the fundamental strategic differentiator of the platform.