TECHNICAL SPECIFICATIONS
Engine & Model Architecture
| ENGINE METRIC | SPECIFICATION | TECHNICAL CAPABILITY |
|---|---|---|
| Inference Engine | TabICLv2 Foundation Model | Tabular In-Context Learning architecture natively understands column structure, formats, and relational values without manual encoding. |
| Hardware Acceleration | WebGPU & WASM Local Runtime | Executes native WebGPU-accelerated tensor kernels directly in your browser. Bypasses backend cloud costs and enables local prediction loops. |
| API Architecture | Model Context Protocol (MCP) | Exposes active model weights as semantic tools. Integrates directly with LLM agents to query prediction ranges via standard MCP schemas. |
| Privacy & Security | Local Sandbox & Hosted Opt-In | Extension runs 100% client-side. For persistent LLM queries, opting into Carla HQ hosts the Spreadsheets KV Cache, no clear-text spreadsheet data ever hits the cloud. |
| Access Control | Enterprise Authentication for all | MCP is authenticated using OAuth 2.1. Available to all users and teams at no additional cost. |
| Explainability | Local SHAP Values | Computes Shapley feature importance attributions natively client-side using WebGPU acceleration. Explains both individual row predictions and global model feature contributions. |
| Inference Endpoints | REST API + MCP Tool Gateways | Supports high-throughput REST API calls for bulk deterministic workflows, alongside MCP integrations that act as decision gates for multi-step autonomous agent loops. |
TECHNOLOGY COMPARISON
TabICLv2 vs. Traditional GBDTs (XGBoost)
Why did we build a tabular foundation model instead of compiling classic gradient-boosted trees?
XGBoost / LightGBM
Traditional GBDT- Heavy Preprocessing Required: Requires manual one-hot encoding for categoricals, scaling for numericals, and explicit handling of missing values or noise before training.
- Hyperparameter Tuning Overhead: Highly sensitive to parameters like learning rate, max depth, and regularization. Demands extensive grid/random search (tuning loops) to prevent overfitting.
- Developer-Only Accessibility: Building, tuning, and deploying models requires a Python, R, or C++ environment, isolating business and financial analysts from direct model creation.
- Static Inference: Models are static once compiled. Adapting to new columns or shifting data schemas requires rebuilding the preprocessing pipeline and retraining from scratch.
TabICLv2 Foundation Model
Tabular In-Context Learning- Zero Preprocessing: Natively understands raw cell formats, categorical labels, text strings, and missing data points out-of-the-box. No manual feature engineering needed.
- Zero Hyperparameter Tuning: Pre-trained across millions of diverse tabular datasets. Generalizes to new schemas instantly using in-context attention weights.
- Democratized AI: Embedded directly into the Google Sheets UI. Allows any spreadsheet user to train, analyze, and inspect models without writing a line of code.
- Dynamic Adaptation: Adapts to row and column edits in real-time. Simply highlight new columns and retrain in-context instantly directly inside the browser sandbox.