Raphaël.D
Building a Confidential AI Platform from 0→1
Building a Confidential AI Platform from 0→1
Building a Confidential AI Platform from 0→1
2025
2025
Enterprise AI Nasdaq-listed
Enterprise AI Nasdaq-listed
AI Design Leader
AI Design Leader
Artificial Intelligence
Artificial Intelligence
Confidential AI & Infrastructure
Confidential AI & Infrastructure
Designing trust in AI: A Confidential Infrastructure Platform
Designing trust in AI: A Confidential Infrastructure Platform
Designed the product UX of a confidential AI platform enabling enterprises to run private AI workloads without exposing sensitive data. From infrastructure architecture to user-facing AI interfaces, built from zero.
Designed the product UX of a confidential AI platform enabling enterprises to run private AI workloads without exposing sensitive data. From infrastructure architecture to user-facing AI interfaces, built from zero.
5
5
Product surfaces designed from 0
Private AI interface, staking, compute provider, user dashboard, and billing. All built from scratch in 4 months.
5
5
Product surfaces designed from 0
Private AI interface, staking, compute provider, user dashboard, and billing. All built from scratch in 4 months.
0→1
0→1
Full product ownership
From research and architecture framing to shipped interfaces, as sole designer across all product surfaces.
0→1
0→1
Full product ownership
From research and architecture framing to shipped interfaces, as sole designer across all product surfaces.
25
+
25
+
Competitors benchmarked
Deep competitive analysis across AI, infrastructure, and incentivization products to define the right design direction.
Nasdaq
Nasdaq
Listed Company
Designing for a publicly traded AI platform. Enterprise-grade standards, verifiable execution, and zero margin for trust failures.
25
+
25
+
AI products benchmarked
Deep competitive analysis across AI, infrastructure, and incentivization products to define the right design direction.
Nasdaq
Nasdaq
Listed Company
Designing for a publicly traded AI platform. Enterprise-grade standards, verifiable execution, and zero margin for trust failures.
25
+
25
+
AI products benchmarked
Deep competitive analysis across AI, infrastructure, and incentivization products to define the right design direction.
Nasdaq
Nasdaq
Listed Company
Designing for a publicly traded AI platform. Enterprise-grade standards, verifiable execution, and zero margin for trust failures.
5
5
Product surfaces designed from 0
Private AI interface, staking, compute provider, user dashboard, and billing. All built from scratch in 4 months.
0→1
0→1
Full product ownership
From research and architecture framing to shipped interfaces, as sole designer across all product surfaces.
Context
AI isn’t limited by performance.
It’s limited by trust.
AI isn’t limited by performance.
It’s limited by trust.
The enterprises that handle the most sensitive, regulated and high-stakes data are the ones AI can't reach yet.
The enterprises that handle the most sensitive, regulated and high-stakes data are the ones AI can't reach yet.
The Problem
Every enterprise AI product today has the same hidden dependency: sensitive data sent to a centralized provider, with no cryptographic proof of how it was handled. For regulated industries like finance, healthcare, legal, government, this isn't a preference. It's a blocker. AI stays out of the workflows that matter most.
The Problem
Every enterprise AI product today has the same hidden dependency: sensitive data sent to a centralized provider, with no cryptographic proof of how it was handled. For regulated industries like finance, healthcare, legal, government, this isn't a preference. It's a blocker. AI stays out of the workflows that matter most.
The Design Challenge
The technical solution exists: Trusted Execution Environments that execute AI inference inside hardware-enforced secure enclaves, with verifiable attestations anchored on-chain. The real challenge: making a system this complex feel like it requires zero trust decisions from the user. That's the design problem OODA hired me to solve.
The Design Challenge
The technical solution exists: Trusted Execution Environments that execute AI inference inside hardware-enforced secure enclaves, with verifiable attestations anchored on-chain. The real challenge: making a system this complex feel like it requires zero trust decisions from the user. That's the design problem OODA hired me to solve.
My role
Led the design of a confidential AI platform. From consumer chat to compute infrastructure.
Led the design of a confidential AI platform. From consumer chat to compute infrastructure.
One designer. Five product surfaces. Every user type from retail consumer to infrastructure provider.
One designer. Five product surfaces. Every user type from retail consumer to infrastructure provider.
01
01
Designing confidential AI for everyone
Led the end-to-end UX of PrivatAI, a consumer-facing conversational AI product built on confidential infrastructure. Designed every interaction from onboarding to prompt, with zero assumption of technical knowledge.
Delivered a ChatGPT-level experience on top of hardware-enforced private inference, making enterprise-grade confidentiality accessible to any user, anywhere.
01
01
Designing confidential AI for everyone
Led the end-to-end UX of PrivatAI, a consumer-facing conversational AI product built on confidential infrastructure. Designed every interaction from onboarding to prompt, with zero assumption of technical knowledge.
Delivered a ChatGPT-level experience on top of hardware-enforced private inference, making enterprise-grade confidentiality accessible to any user, anywhere.
01
01
Designing confidential AI for everyone
Led the end-to-end UX of PrivatAI, a consumer-facing conversational AI product built on confidential infrastructure. Designed every interaction from onboarding to prompt, with zero assumption of technical knowledge.
Delivered a ChatGPT-level experience on top of hardware-enforced private inference, making enterprise-grade confidentiality accessible to any user, anywhere.
02
02
Replacing model selection with human context
Decided to abstract LLM selection behind a personality layer. Users choose a Writing Pal, Therapist Pal, or Health Pal instead of a model name. Defined each Pal's behavior, tone, and use case scope.
Reduced cognitive load at the point of highest friction: the moment a user decides whether AI is useful to them. Made the product feel personal before it felt technical.
02
02
Replacing model selection with human context
Decided to abstract LLM selection behind a personality layer. Users choose a Writing Pal, Therapist Pal, or Health Pal instead of a model name. Defined each Pal's behavior, tone, and use case scope.
Reduced cognitive load at the point of highest friction: the moment a user decides whether AI is useful to them. Made the product feel personal before it felt technical.
02
02
Replacing model selection with human context
Decided to abstract LLM selection behind a personality layer. Users choose a Writing Pal, Therapist Pal, or Health Pal instead of a model name. Defined each Pal's behavior, tone, and use case scope.
Reduced cognitive load at the point of highest friction: the moment a user decides whether AI is useful to them. Made the product feel personal before it felt technical.
03
03
Making Web3 invisible by design
Pushed back on the default assumption that users would connect an external wallet. Decided instead to auto-provision a self-hosted wallet at account creation. The entire token and staking layer operates without the user ever seeing a wallet interface.
Eliminated the single biggest onboarding barrier for non-crypto users. The product inherits all the security guarantees of a Web3 infrastructure while feeling like a standard SaaS signup.
03
03
Making Web3 invisible by design
Pushed back on the default assumption that users would connect an external wallet. Decided instead to auto-provision a self-hosted wallet at account creation. The entire token and staking layer operates without the user ever seeing a wallet interface.
Eliminated the single biggest onboarding barrier for non-crypto users. The product inherits all the security guarantees of a Web3 infrastructure while feeling like a standard SaaS signup.
03
03
Making Web3 invisible by design
Pushed back on the default assumption that users would connect an external wallet. Decided instead to auto-provision a self-hosted wallet at account creation. The entire token and staking layer operates without the user ever seeing a wallet interface.
Eliminated the single biggest onboarding barrier for non-crypto users. The product inherits all the security guarantees of a Web3 infrastructure while feeling like a standard SaaS signup.
04
04
Turning verification into a human proof
Designed the Explorer as an audit trail, not a block explorer. Every AI interaction generates a cryptographic attestation anchored on-chain. I decided to surface this as "proof your conversation was private" rather than as a transaction history.
Transformed a technically complex Web3 primitive into a trust signal that any enterprise compliance officer could understand and point to during an audit.
04
04
Turning verification into a human proof
Designed the Explorer as an audit trail, not a block explorer. Every AI interaction generates a cryptographic attestation anchored on-chain. I decided to surface this as "proof your conversation was private" rather than as a transaction history.
Transformed a technically complex Web3 primitive into a trust signal that any enterprise compliance officer could understand and point to during an audit.
04
04
Turning verification into a human proof
Designed the Explorer as an audit trail, not a block explorer. Every AI interaction generates a cryptographic attestation anchored on-chain. I decided to surface this as "proof your conversation was private" rather than as a transaction history.
Transformed a technically complex Web3 primitive into a trust signal that any enterprise compliance officer could understand and point to during an audit.
05
05
One designer across five product surfaces
Operated as sole designer across PrivatAI, staking dashboard, compute provider interface, user account management, and Explorer. Each surface served a fundamentally different user type with different mental models and expectations.
Maintained design coherence across a product where a retail consumer, an enterprise buyer, and an infrastructure provider all interact with the same underlying system, without any of them needing to understand each other's layer.
05
05
One designer across five product surfaces
Operated as sole designer across PrivatAI, staking dashboard, compute provider interface, user account management, and Explorer. Each surface served a fundamentally different user type with different mental models and expectations.
Maintained design coherence across a product where a retail consumer, an enterprise buyer, and an infrastructure provider all interact with the same underlying system, without any of them needing to understand each other's layer.
05
05
One designer across five product surfaces
Operated as sole designer across PrivatAI, staking dashboard, compute provider interface, user account management, and Explorer. Each surface served a fundamentally different user type with different mental models and expectations.
Maintained design coherence across a product where a retail consumer, an enterprise buyer, and an infrastructure provider all interact with the same underlying system, without any of them needing to understand each other's layer.
Deep Dive
Private AI
Private AI
Designing familiarity on top of complexity.
Designing familiarity on top of complexity.
The Problem
Enterprise-grade confidential AI infrastructure means nothing to a user who just wants to ask a question privately. The risk was building something technically impressive but experientially alienating, a product that signals "this is complicated" before it signals "this is useful."
The Problem
Enterprise-grade confidential AI infrastructure means nothing to a user who just wants to ask a question privately. The risk was building something technically impressive but experientially alienating, a product that signals "this is complicated" before it signals "this is useful."
The Decision
I designed the entire conversational experience to feel like a product users already knew, with familiar interface patterns, natural onboarding, and zero visible infrastructure. The "Hello, Stranger" entry point was deliberate: warm, human, no technical framing. Confidentiality happens underneath, invisibly.
The Decision
I designed the entire conversational experience to feel like a product users already knew, with familiar interface patterns, natural onboarding, and zero visible infrastructure. The "Hello, Stranger" entry point was deliberate: warm, human, no technical framing. Confidentiality happens underneath, invisibly.
Outcome
A consumer AI product that inherits hardware-enforced privacy guarantees without asking users to understand what that means. The complexity is real. The experience of it isn't.
Outcome
A consumer AI product that inherits hardware-enforced privacy guarantees without asking users to understand what that means. The complexity is real. The experience of it isn't.
Deep Dive
AI Pals
AI Pals
Replacing a technical choice with a human one.
Replacing a technical choice with a human one.
The Problem
Exposing LLM selection to users creates immediate cognitive friction. Most people don't know what GPT-4o vs Claude vs Mistral means for their use case. Asking them to choose a model is asking them to make a decision they're not equipped to make.
The Problem
Exposing LLM selection to users creates immediate cognitive friction. Most people don't know what GPT-4o vs Claude vs Mistral means for their use case. Asking them to choose a model is asking them to make a decision they're not equipped to make.
The Decision
I removed model selection from the primary flow entirely and replaced it with use-case personas: Research Pal, Legal Pal, Writing Pal. Each Pal has a defined behavior, tone, and scope. The user picks their context, not their infrastructure.
The Decision
I removed model selection from the primary flow entirely and replaced it with use-case personas: Research Pal, Legal Pal, Writing Pal. Each Pal has a defined behavior, tone, and scope. The user picks their context, not their infrastructure.
Outcome
Reduced the highest-friction moment in AI onboarding to a single, intuitive choice. The product feels personal immediately, before it feels technical. Model selection still exists for power users, but it's never the first question.
Outcome
Reduced the highest-friction moment in AI onboarding to a single, intuitive choice. The product feels personal immediately, before it feels technical. Model selection still exists for power users, but it's never the first question.
Deep Dive
Explorer
Explorer
Turning cryptographic proof into human trust.
Turning cryptographic proof into human trust.
The Problem
Every PrivatAI interaction generates a cryptographic attestation anchored on-chain, provable evidence that the computation happened inside a TEE and that no data was exposed. This is a powerful enterprise trust signal. It's also completely inaccessible to anyone who doesn't understand blockchain.
The Problem
Every PrivatAI interaction generates a cryptographic attestation anchored on-chain, provable evidence that the computation happened inside a TEE and that no data was exposed. This is a powerful enterprise trust signal. It's also completely inaccessible to anyone who doesn't understand blockchain.
The Decision
I designed the Explorer not as a block explorer but as an audit trail. The framing shifted from "here are your on-chain transactions" to "here is proof that your conversation was private." Each entry shows what was processed, when, and the verification status, in language a compliance officer can read, not a Web3 developer.
The Decision
I designed the Explorer not as a block explorer but as an audit trail. The framing shifted from "here are your on-chain transactions" to "here is proof that your conversation was private." Each entry shows what was processed, when, and the verification status, in language a compliance officer can read, not a Web3 developer.
Outcome
Transformed a Web3-native primitive into an enterprise trust layer. A feature that could have been invisible or confusing became a differentiating proof point, something an enterprise buyer can show their legal or compliance team without explanation.
Outcome
Transformed a Web3-native primitive into an enterprise trust layer. A feature that could have been invisible or confusing became a differentiating proof point, something an enterprise buyer can show their legal or compliance team without explanation.
Project Timeline
From Research to Mainnet
From Research to Mainnet
Five phases. One design principle throughout: make the infrastructure disappear, keep the trust visible.
Five phases. One design principle throughout: make the infrastructure disappear, keep the trust visible.
Phase 0
Research & Foundation
Mapping a system nobody had designed before:
Benchmarked 25+ AI and infrastructure products.
Structured the product architecture across 5 surfaces.
Defined the core design principle : infrastructure complexity stays invisible, trust signals stay visible.
Phase 0
Research & Foundation
Mapping a system nobody had designed before:
Benchmarked 25+ AI and infrastructure products.
Structured the product architecture across 5 surfaces.
Defined the core design principle : infrastructure complexity stays invisible, trust signals stay visible.
Phase 1
Private AI & Consumer Experience
Designing the front door of a confidential AI platform:
Built the full PrivatAI conversational experience: onboarding, AI Pals, model selection, and the Explorer as audit trail.
First product surface shipped.
Phase 1
Private AI & Consumer Experience
Designing the front door of a confidential AI platform:
Built the full PrivatAI conversational experience: onboarding, AI Pals, model selection, and the Explorer as audit trail.
First product surface shipped.
Phase 2
Infrastructure Layer
Translating Web3 mechanics into familiar product patterns:
Designed staking dashboard, wallet abstraction, and user asset management.
Made token mechanics feel like a SaaS subscription tier, not a crypto product.
Phase 2
Infrastructure Layer
Translating Web3 mechanics into familiar product patterns:
Designed staking dashboard, wallet abstraction, and user asset management.
Made token mechanics feel like a SaaS subscription tier, not a crypto product.
Phase 3
Provider Ecosystem
Extending the same design language to a completely different user:
Explored compute provider dashboard, applying the same abstraction principles to infrastructure operators.
Different user, same philosophy.
Phase 3
Provider Ecosystem
Extending the same design language to a completely different user:
Explored compute provider dashboard, applying the same abstraction principles to infrastructure operators.
Different user, same philosophy.
Phase 4
Mainnet Preparation
Scaling design decisions to production standards:
Design system consolidated.
Enterprise onboarding flows refined.
Product ready for Q2 2026 mainnet launch.
Phase 4
Mainnet Preparation
Scaling design decisions to production standards:
Design system consolidated.
Enterprise onboarding flows refined.
Product ready for Q2 2026 mainnet launch.
AI-Embedded Workflow
AI wasn't a shortcut.
It was a force multiplier at every stage.
AI wasn't a shortcut.
It was a force multiplier at every stage.
Three moments where AI changed not just how fast I worked, but what I was able to do alone.
Three moments where AI changed not just how fast I worked, but what I was able to do alone.
Research & Competitive Intelligence
Benchmarking 25+ products manually would have taken weeks with no guarantee of finding the right patterns. I used AI to map and cluster competitors by category: staking mechanics, conversational AI, enterprise dashboards, surfacing UX patterns from outside the obvious space. Weeks of research compressed into days.
Vibe Coding & Rapid Exploration
As the sole designer across 5 surfaces, I couldn't afford long exploration cycles in Figma. I used Lovable and Figr to vibe-code UI directions directly, generating components and flows to react to immediately. What worked got refined. What didn't got discarded in minutes. Exploration cycles cut from a week to a day.
Design System & Feedback Loop
Building a design system solo while shipping product simultaneously is a contradiction. I used AI to generate variants, document decisions, and maintain consistency guidelines, then fed those guidelines back into the prototyping tools to generate new features already aligned with the system. One designer. Five surfaces. No inconsistency.
Research & Competitive Intelligence
Benchmarking 25+ products manually would have taken weeks with no guarantee of finding the right patterns. I used AI to map and cluster competitors by category: staking mechanics, conversational AI, enterprise dashboards, surfacing UX patterns from outside the obvious space. Weeks of research compressed into days.
Vibe Coding & Rapid Exploration
As the sole designer across 5 surfaces, I couldn't afford long exploration cycles in Figma. I used Lovable and Figr to vibe-code UI directions directly, generating components and flows to react to immediately. What worked got refined. What didn't got discarded in minutes. Exploration cycles cut from a week to a day.
Vibe Coding & Rapid Exploration
As the sole designer across 5 surfaces, I couldn't afford long exploration cycles in Figma. I used Lovable and Figr to vibe-code UI directions directly, generating components and flows to react to immediately. What worked got refined. What didn't got discarded in minutes. Exploration cycles cut from a week to a day.
Design System & Feedback Loop
Building a design system solo while shipping product simultaneously is a contradiction. I used AI to generate variants, document decisions, and maintain consistency guidelines, then fed those guidelines back into the prototyping tools to generate new features already aligned with the system. One designer. Five surfaces. No inconsistency.
Design System & Feedback Loop
Building a design system solo while shipping product simultaneously is a contradiction. I used AI to generate variants, document decisions, and maintain consistency guidelines, then fed those guidelines back into the prototyping tools to generate new features already aligned with the system. One designer. Five surfaces. No inconsistency.
Key Takeaways
What designing a confidential AI platform taught me about trust, complexity, and product thinking.
What designing a confidential AI platform taught me about trust, complexity, and product thinking.
Three decisions that shaped every screen, every flow, and every trust signal on the platform.
Three decisions that shaped every screen, every flow, and every trust signal on the platform.
Trust is an architectural decision, not a design layer.
The most important design work happens before any screen exists.
One designer can move faster than a team, if the system is right.
Trust is an architectural decision, not a design layer.
The most important design work happens before any screen exists.
One designer can move faster than a team, if the system is right.
Raphaël.D
AI Design Leader
© 2026 Raphaël.D. All rights reserved.
Made from scratch in Framer with 💜
Raphaël.D
AI Design Leader
© 2026 Raphaël.D. All rights reserved.
Made from scratch in Framer with 💜
Raphaël.D
AI Design Leader
© 2026 Raphaël.D. All rights reserved.
Made from scratch in Framer with 💜