SuperWisdom.ai
Zero-knowledge AI. Built in three days.
SuperWisdom.ai needed a fiduciary-grade confidential AI platform that could go to market immediately. We built the entire product from zero to production in a 3-day sprint: 59 commits, 23 source files, 3,765 lines of TypeScript. Real-time streaming AI chat with word-by-word markdown rendering. Per-customer AWS isolation with dedicated encryption keys. Ephemeral conversations that exist only in volatile memory and vanish on browser close. Seven AWS services orchestrated into a zero-knowledge architecture where not even SuperWisdom can access customer data.
Overview
Challenge
A federal judge's ruling in the Heppner case established that attorney-client privilege can be waived through standard AI platform usage. Law firms, family offices, fiduciaries, and trustees needed access to frontier AI models but couldn't trust enterprise contracts that offer only contractual promises, not architectural guarantees. Palantir serves Fortune 500 companies. Internal IT teams handle large enterprises. But no mid-market solution existed for organizations that depend on confidentiality and need genuine zero-knowledge AI infrastructure, not just a terms-of-service checkbox.
Solution
We built SuperWisdom.ai v1 as a complete 0-to-1 platform in a 3-day sprint. A Next.js 15 frontend with React 19 and real-time SSE streaming that renders AI responses word-by-word with animated markdown. A dual inference architecture: Lambda streaming functions with Function URLs as the primary path for true server-to-browser SSE, with a Next.js API fallback route. AWS Bedrock for multi-model Claude inference (Sonnet 4.5, Sonnet 3.7, Opus 4.1). Cognito OAuth with JWT tokens stored in httpOnly cookies to prevent XSS theft. S3 file storage with AES-256-GCM envelope encryption where each conversation gets its own data encryption key via KMS. Textract for PDF OCR and text extraction. Bedrock Guardrails for configurable content moderation. A fully installable PWA for desktop and mobile. All deployed on Amplify with 15-minute compute timeouts for long streaming operations.
Outcome
SuperWisdom launched to market with enterprise-grade security from day one. The zero-knowledge architecture became the core of their sales narrative and competitive positioning. The platform was designed for extensibility, and SuperWisdom's team subsequently expanded it with web search, SageMaker models, stored conversations, memory features, WisdomDrive storage, export capabilities, and subscriber issue reporting. The product now serves paying customers at $299/month per user with every subscriber treated as enterprise. SuperWisdom is targeting FedRAMP High and DoD IL5/IL6 compliance and is leading the SOC Zero nonprofit initiative for zero-trust/zero-knowledge standards.
The Problem
The confidentiality gap in AI
Every major AI provider offers enterprise contracts. OpenAI, Anthropic, Google all promise they won't train on your data, won't retain it beyond processing, and will protect it under their security frameworks. For most businesses, that's sufficient.
For fiduciaries, it's not. Attorneys have a legal obligation to protect client privilege. Trustees have fiduciary duties that extend to information security. Family offices managing billions in assets can't accept 'we promise not to look' as a security architecture. The Heppner ruling made this concrete: a federal judge found that using standard AI platforms could waive attorney-client privilege. The contractual guarantee wasn't enough because the data still transited through infrastructure the firm didn't control.
The market had two extremes. Fortune 500 companies could deploy Palantir or build internal solutions with dedicated engineering teams. Everyone else was stuck with enterprise AI subscriptions that offered contractual comfort but not architectural certainty. No mid-market product existed that gave a 30-person law firm the same zero-knowledge infrastructure that a defense contractor would demand.
That's the gap SuperWisdom set out to fill. Not better contracts. Better architecture.
The Architecture
Zero-knowledge by design, not by policy
The security model starts with isolation. Every SuperWisdom customer gets their own dedicated AWS sub-account. Their own Cognito identity pool. Their own DynamoDB tables. Their own S3 buckets. Their own KMS encryption keys. Their own Bedrock guardrail configuration. No shared infrastructure. No multi-tenant databases. If Customer A's environment were entirely deleted, Customer B would never know.
Encryption goes deeper than standard at-rest protection. Each conversation gets its own data encryption key via AES-256-GCM envelope encryption managed by KMS. File attachments uploaded to S3 are encrypted with customer-specific keys. The separation isn't just logical, it's cryptographic. Even SuperWisdom's own engineering team cannot decrypt customer data because the keys live in the customer's isolated AWS account.
The default mode is ephemeral. Conversations exist only in the browser's volatile memory. Close the tab, and the conversation is gone. There's no server-side session storage, no conversation history database, no backup. This isn't a feature toggle or a premium tier. It's the foundational architecture. The data never exists in a form that could be subpoenaed, breached, or accessed by SuperWisdom employees.
Authentication uses Cognito OAuth with JWT tokens stored exclusively in httpOnly cookies, which prevents JavaScript from accessing them and eliminates an entire class of XSS attacks. Token refresh happens automatically on 401 responses. Full JWT signature verification runs on every request.
This is the difference between a contractual promise and an architectural guarantee. You can't leak what you never stored. You can't access what you can't decrypt. You can't breach an isolated environment you don't share.
The Transformation
What we rebuilt
Real-Time Streaming Chat
ChatGPT-class streaming AI interface with dual inference paths. Primary: Lambda streaming function with Function URL for true server-to-browser SSE. Fallback: Next.js API route. Word-by-word markdown rendering with 30ms animation, GitHub-flavored markdown support, and SSE heartbeat pings. Model selection locked after first message to maintain conversation coherence.
Zero-Knowledge Architecture
Conversations exist only in volatile browser memory and vanish on close. No server-side storage, no history database, no backups. AES-256-GCM envelope encryption with per-conversation data keys via KMS. Not even SuperWisdom can access customer data. The architecture makes breach, subpoena, and insider access structurally impossible.
Per-Customer Isolation
Every customer gets a dedicated AWS sub-account with isolated Cognito pool, DynamoDB, S3, KMS keys, and Bedrock guardrails. No shared infrastructure. No multi-tenant databases. IAM least privilege with separate users for Bedrock vs. S3/Textract. Deletion of one customer environment has zero impact on any other.
Multi-Model AI Inference
AWS Bedrock integration supporting Claude Sonnet 4.5, Sonnet 3.7, and Opus 4.1. Model selection per conversation with lock-after-first-message. Configurable content moderation via Bedrock Guardrails. 15-minute compute timeout for long-running inference operations. Graceful fallback between Lambda streaming and API route inference paths.
Secure File Processing
PNG, JPEG, and PDF uploads with client-side Sharp compression before transfer. 10MB per file, 8 files max per message. PDFs processed through AWS Textract for OCR text extraction. All files encrypted with customer-specific KMS keys in isolated S3 buckets. File context injected into AI prompts for document-aware conversations.
Progressive Web App
Fully installable on desktop and mobile via PWA manifest. Responsive design optimized for legal and financial professionals who work across devices. Offline-capable shell with network-dependent AI inference. Custom favicon and app icons (192px, 512px) for native-feeling installation.
Architecture
The new foundation
A security-first serverless architecture where every layer is designed around zero-knowledge principles. Dual inference paths provide streaming resilience. Per-customer AWS sub-accounts ensure complete isolation. Envelope encryption with per-conversation keys means data is cryptographically separated even within a single customer's environment. The entire platform was built to be extensible, and SuperWisdom's team has since added web search, SageMaker models, stored conversations, and file management on top of the original architecture.
Next.js 15 Frontend
Streaming chat UI with word-by-word markdown rendering
Lambda Streaming
Primary inference path for real-time AI streaming
Next.js API Fallback
Secondary inference path for streaming resilience
AWS Bedrock
Multi-model AI inference with guardrail moderation
AWS Cognito
Authentication with XSS-resistant token storage
AWS S3 + KMS
Encrypted file storage with envelope encryption
AWS Textract
PDF text extraction for document-aware AI conversations
AWS Amplify
Production hosting with long-running compute support
Technology Stack
The modern toolkit
The Result
See what emerged
AI Chat Interface
A ChatGPT-class streaming interface built for confidential work. Real-time SSE streaming with word-by-word markdown rendering, model selection, and conversation controls. Dark brown/tan theme designed for professionals who spend hours in the interface.
Streaming Responses
Word-by-word markdown rendering with 30ms animation, code blocks, lists, and formatting
Model Selection
Choose between Claude Sonnet 4.5, Sonnet 3.7, and Opus 4.1 with model lock after first message
Ephemeral Mode
Conversations exist only in volatile memory and vanish completely on browser close
Secure File Processing
Upload documents and images for AI-assisted analysis. Client-side compression, encrypted S3 storage, and Textract OCR for PDF text extraction. All files encrypted with customer-specific KMS keys.
File Upload
PNG, JPEG, and PDF uploads with client-side Sharp compression. 10MB per file, 8 files per message
PDF OCR
AWS Textract extracts text from PDFs and scanned documents for AI context injection
Encrypted Storage
AES-256-GCM envelope encryption with per-conversation data keys in isolated S3 buckets
Security Architecture
Every security decision was architectural, not contractual. Per-customer AWS isolation, per-conversation encryption keys, ephemeral-by-default conversations, and httpOnly JWT cookies. Designed to exceed SOC 2 and target FedRAMP High.
Customer Isolation
Dedicated AWS sub-account per customer with isolated Cognito, DynamoDB, S3, KMS, and Bedrock guardrails
Envelope Encryption
AES-256-GCM with per-conversation data encryption keys managed by customer-specific KMS
Authentication
Cognito OAuth with JWT in httpOnly cookies, automatic refresh on 401, full signature verification
Content Guardrails
AWS Bedrock Guardrails for configurable content sensitivity per customer environment
Progressive Web App
Fully installable on desktop and mobile with native-feeling experience. Responsive design for legal and financial professionals who work across devices.
App Installation
PWA manifest with custom icons (192px, 512px) for desktop and mobile installation
Responsive Design
Optimized layouts for desktop workstations, tablets, and mobile devices
Ready to build something
extraordinary?
Let's discuss how we can bring your vision to life.