An AI copilot that lives inside Excel, where most real financial modeling actually happens. You tell it what you want in plain English; it plans, shows you the plan, then builds and edits the live spreadsheet — with a verification pass that re-checks every formula it touched, because in a financial model one wrong cell propagates silently.
A large share of real financial modeling happens in Excel: hundreds of interlinked formulas across sheets, house formatting standards, inherited workbooks, dense regulation to cross-check. Generic chatbots can't help much because they can't see or touch the workbook.
An add-in exposes about twenty spreadsheet operations to a multi-agent backend: an intake agent asks targeted clarifying questions, a planner produces a phased plan the user must approve before any cell is touched, an executor streams the actual edits into the live workbook, and a verifier re-reads every modified range hunting for formula errors, hardcoded constants, and cross-sheet inconsistencies. Every action appears as a card, so the user sees exactly what was done. It can also convert a workbook into a runnable Python model and validate the generated code by reconciling its outputs against the original.
Agentic orchestration with client-executed tools; retrieval over a curated knowledge library — US reserve regulation summarized in plain terms, firm modeling standards — with cited sources; multi-provider model routing with automatic fallback; and an eval harness with regression suites and cost telemetry, so agent quality is measured, not anecdotal.
TypeScript, React, Office.js; Node + Express; Python, FastAPI, OpenAI Agents SDK, LiteLLM (OpenAI / Azure / Anthropic / Gemini), ChromaDB; Playwright-driven in-Excel testing.