Overview
Key Features
SpreadsheetBench Verified (400) Milestone: Scored 82.75% (passing 331/400 curated tasks) on the official independently evaluated SpreadsheetBench V1 Verified leaderboard, outperforming raw frontier models and competing directly alongside enterprise labs.
SpreadsheetBench Global Top 10: Verified #10 ranking on the official SpreadsheetBench V1 - Full (912 tasks across 2,729 sheets) evaluation, beating frontier AI baselines on complex relational grids.
Deterministic Client Sandbox: Executes JavaScript transformations in an isolated browser runtime, never probabilistically guessing math, and catching typos via self-healing code loops.
ReAct+ Agent Engine: Multi-turn self-healing reflection (openpyxl & pandas) for serverless spreadsheet automation, document extraction, and sheet auditing.
Amber Preview & Rollback Snapshots: Inspect cell-level changes with glowing amber diff highlights before applying, backed by a full multi-turn undo/redo state snapshot engine.
8 Interactive Visuals & Executive Dashboards: Renders Line, Area, Composed, Horizontal Bar, Radar, Scatter, Bar, and Pie charts directly in-thread or compiled into custom multi-grid BI dashboards with elite corporate advisory briefs.
Dual-Stream LLMOps & Observability: Full execution tracing to Langfuse Cloud and native Convex trace engine waterfalls with latency percentiles, token burn tracking, and an operation-aware credit deduction matrix.
Autonomous Web-to-Sheet & arXiv Ingestion: Live web search and deep scraping via Firecrawl/LangSearch to enrich sheet columns, plus direct arXiv API ingestion of scholarly abstracts and PDFs.
Byte-Safe Volume Protection & Multi-Format Ingestion: Sub-second AG Grid operations on 10,000+ row files with auto-flattening support for Excel (.xlsx, .xls), CSV, and nested JSON structures.