RnD Ventures
AI OS
Designing systems
powered by
conversational AI
and a "Liquid Interface"Designing a system powered by conversational AI and a "Liquid Interface" for the Post-Web Era
01
The Context: Surviving in the "AI Search Era"
1.1 The Role: Redesigning digital identity (Team Effort)
Initially, I participated as UX Lead on a multidisciplinary team to redesign the corporate website for rnd.ventures. The goal was to modernize the RnD Ventures website. We delivered a functional, aesthetic, and modern site.
However, before launch, I had already identified a strategic vulnerability while developing the project Multi-Agent AEO & GEO Assistant: The "Website" model is dying.
1.2 The Transformation of the Digital Ecosystem
Based on recent market analysis, I proposed to stakeholders that a static website was no longer sufficient. Tools like ChatGPT, Claude, or Perplexity are "devouring" content to train their models, eliminating the need for users to visit the original source.
The data is alarming and justifies the creation of RnD Ventures AI OS:
The Bot Invasion
Scraping ratio vs. real visits — Source: Matthew Prince, CEO of Cloudflare
6 months ago
Now
Generative AI
Por cada 1 humano que visita tu web, hay 60.000 bots extrayendo tu contenido.
1.3 The Paradigm Shift
The Conclusion
If users no longer navigate but ask AIs, RnD Ventures could not limit itself to having a "brochure" website.
What I proposed was a conversational AI operating system with a liquid interface — a primary point of contact capable of controlling the narrative and offering instant value without friction.
02
Pain Points: Why Traditional Web Fails in VC
My research identified 5 critical friction points that drive users to abandon traditional websites:
Inefficient Navigation
A founder spends 45-60 minutes reading theses, searching for information, and finding the steps to follow.
Instant response.
Cognitive Overload
Websites saturated with legal jargon. Mental fatigue from excessive text.
Explicaciones simplificadas y visuales bajo demanda.
Fragmented Information
The portfolio is on one page, the team on another, the application on another, etc.
Centralization. The agent brings information to the user, not the other way around.
Language Barriers
rnd.ventures speaks English, but talent is global. 85% of technical content excludes stakeholders from many countries.
Native multilingual.
03
The Solution: RnD Ventures AI OS
I designed RnD Ventures AI OS as a Conversational Operating System — an intelligence layer whose goal is to give RnD Ventures the ability to deliver a personalized response to every stakeholder. It is not a chatbot; it is a new way of interacting with rnd.ventures in the post-web era.
3.1 UX Concept: "Liquid Interface" (GenUI)
The system delivers responses as a continuous stream or interactive storytelling (similar to a feed), where each response can combine text, visual components, or assist chips.
3.2 Transparency: "The Run Log"
To mitigate distrust in the AI "Black Box", I designed a collapsible panel ("Run Log").
Functionality
Allows users to see the reasoning trace of the AI in real time.
Impact
Users can see which files the system consults, generating trust.
04
Engineering & Security: Technical Audit
Since the system can handle sensitive information, the MVP was not launched until it passed a strict battery of security and compliance audits (Red Teaming & FAST Framework).
4.1 4-Layer Defense Architecture
Based on the security report from 02/14/2026.
Layer 0: DLP Shield
Red ShieldScanning with Google Cloud DLP (Europe Region). Blocks PII before processing.
Layer 1: Deterministic Routing
YAML FlowsStrict YAML files for critical flows. Zero hallucinations.
Layer 2: RAG Seguro
Verified RAGInjection of validated knowledge with SHA-256 checksums.
Layer 3: Generative Core
Gemini 3Gemini 3 Flash con System Instructions inmutables.
4.1.5 Anatomy of an Interaction
The following diagram shows the complete flow of a request through the 6 phases of the system, from ingestion to audit.
Architecture Diagram
6 fases · 18 nodos
Ingestion — Next.js
The user sends a message from the chat interface. Next.js captures it via a Server Action (sendMessage), keeping API keys exclusively on the server.
Zod validates the payload schema: if it fails, the error returns to the UI immediately without crossing any subsequent security layer.
Layer 0 — Red Shield
First line of defense. A local regex scanner detects basic PII (national ID, IBAN, emails) with no network latency.
If there is ambiguity, the payload goes to Google Cloud DLP—anchored in Europe-West1 for GDPR compliance—which applies more precise ML models. If PII is detected at any point, the request is blocked and redacted before advancing to the LLM.
Hybrid Routing
The Intent Router classifies the message intent. If it matches a critical flow defined in .yaml files (questions about investments, applications, team), the response is 100% deterministic: served directly from YAML without invoking the LLM, eliminating hallucination risk in high-risk flows. Only unmapped queries advance to Phase 3.
Secure RAG — SHA-256 Verification
Verified knowledge .md files are loaded. Before injecting them as context, an SHA-256 checksum is computed for each file and compared to the hash registered at deployment.
If the hash does not match (possible knowledge manipulation attack), the request is blocked. If integrity is verified, the context is injected into the System Prompt and sent to Gemini 1.5 Flash with temperature 0.2 to maximize precision.
Gen-UI Rendering — Liquid Interface
The Parser analyzes the model response. If it detects structured JSON (agreed upon in the System Prompt), React renders interactive components: metric cards, carousels, charts.
If the response is plain text, it is rendered directly. This bifurcation is the core of the Liquid Interface: the same architecture delivers complex UI or simple text based on detected intent.
System Online. I am the RnD Ventures Intelligence Layer. Which aspect of the ecosystem would you like to explore?
Metrics are for ecosystem health monitoring only. Information provided for informational purposes only. This AI does not provide financial advice.
FAST Audit — Immutable Traceability
In parallel to the entire flow (dashed line in the diagram), the Server Action signs each session with SHA-256 and writes an immutable log to Firestore (europe-west1, 90-day TTL).
This log feeds the Run Log visible in the UI, allowing users to see the complete reasoning trace. Complies with Art. 12 of EU AI Act (Traceability) and the Transparency principle of the FAST Framework (10/10).
4.1.6 Request Lifecycle
Time is the invisible variable that defines the experience.
This diagram breaks down the complete lifecycle of a request—from when the user presses "Send" until they receive the rendered response—showing the duration of each phase and where latency accumulates.
Optimizing each segment is what separates a slow chatbot from a system that feels instant.
Request Lifecycle
latencia · fases · tiempo
4.2 Regulatory Compliance
Regulatory Compliance
Based on the Technical-Legal Compliance Audit Report (February 2026)
The system has passed the complete architecture audit under the GDPR (EU 2016/679) and EU AI Act (EU 2024/1689) frameworks.
Privacy-by-Design Architecture
The 4-layer architecture ensures that no sensitive data reaches the LLM model without prior sanitization.
FAST Framework Scores
Technical-Legal Compliance Audit Report · February 2026
Fairness
Art. 14The system operates under a "Non-decisional" framework; the AI has no authority for binding commitments verified ToS Sec. 5.1
Accountability
Art. 12Dual-log strategy: conversation_audit + quality_debug
Security & GDPR
Data SovereigntyBlocks ID, IBAN, Emails before reaching Gemini
Transparency
Art. 50"You are The Analyst — an AI assistant"The system proactively declares its artificial nature
Disclaimers about the non-deterministic and probabilistic nature of the integrated LLMs in the Terms of Service.
05
Execution Results and Roadmap
The RnD Ventures AI OS has evolved from a proof of concept to a Production-Ready MVP.
Production (Private Alpha)
February 2026
Performance
Latency
Average end-to-end latency. The 4-layer architecture does not compromise speed.
Security
Vulnerabilities
Technical audit confirmed under European compliance frameworks.
Resilience
Layer
Proven ability to handle prompt injections and block sensitive PII.
Next Steps
Post-Audit Roadmap
Based on the findings Advisory from the audit report, the immediate roadmap focuses on:
GDPR Operationalization
AIA-03Development of an internal administration panel to facilitate ARCO rights management.
Policy Refinement
AIA-02Update of legal documentation and refinement of data retention policies.
"This project demonstrates that regulatory security is not a brake, but an enabler."
"We have managed to merge a modern conversational UX with rigorous compliance engineering, creating a system capable of surviving and scaling in an increasingly regulated and automated digital ecosystem."
RnD Ventures AI OS
Frequently Asked Questions about AI OS
Everything you need to know about how an AI Operating System can transform your business's digital presence.
Let's talk?
Are you looking for a UX designer for Artificial Intelligence who can help broaden perspectives and reduce biases in conversational AI?
Fill out the form below or, if you prefer, write to me directly at info@josegalan.dev and let's see how we can work together.