Advantora Insights Security Architecture

Most AI data analysts force a difficult compromise: upload your sensitive corporate data to a cloud analysis workspace, or miss out on AI-assisted reporting. Advantora Insights was engineered differently. By combining secure persistent storage with local browser compute for spreadsheets, we deliver boardroom-ready AI analysis with clearer data-processing boundaries.

1. Secure Persistent Workspaces (Storage)

To ensure you never lose your analysis, your uploaded files and generated reports are saved to a secure cloud backend.

• Encryption at Rest: All files are stored in S3-backed storage with industry-standard AES-256 encryption.
• Row-Level Security (RLS): Advantora utilizes strict Row-Level Security at the database core. Your workspace is cryptographically isolated. No other userand no unauthorized systemcan query or access your storage folders.

2. The Bifurcated Compute Engine (Processing)

Not all data carries the same risk. Advantora routes your data through different compute engines based on its format to maximize both privacy and performance.

A. Structured Data (CSVs & Spreadsheets) → Tokenization & Local WASM

Your spreadsheets contain your most sensitive PII and financial data. When you analyze a CSV, Advantora pulls the file from your encrypted storage directly into your local browser's RAM using DuckDB WebAssembly.

Your structured spreadsheet queries run locally. The AI helps interpret intent and draft SQL, while the row-level calculation is executed in the browser. Depending on the workflow, schema, samples, aggregated results, prompts, or document content may be sent through secure AI API processing.

B. Unstructured Data (PDFs) → Enterprise Edge Routing

Because local browser databases cannot query a 30-page legal PDF, unstructured text must be sent to an LLM for summarization and cross-referencing.

For these tasks, Advantora routes data through secure edge functions to AI API endpoints. API providers such as OpenAI state that business/API data is not used to train or improve models by default, though retention and abuse-monitoring settings can vary by provider, endpoint, and account configuration.

3. Determinative Local Execution (Accuracy)

Standard public LLMs suffer from mathematical "hallucinations" because they try to guess numbers based on text prediction. Advantora solves this by using the AI strictly as a query translation layer.

The LLM translates your plain-English prompt into SQL. That SQL is then executed locally via DuckDB, making calculations inspectable and repeatable before the final presentation is generated.