Corda AI

We built an AI-powered platform that automates conference scheduling and meeting coordination for capital markets teams at investment banks. It reads the conference inbox, extracts issuer availability and investor meeting requests, detects conflicts, and coordinates meetings between bankers and Sales. It also optimizes banker coverage and meeting-room assignment based on availability and preferences, and builds the conference schedule automatically, with every change approved by a banker in one click. Models run on Nicrron, and every agent step is traced in PRISM.

We started building this because one of our founders works in capital markets at a major investment bank and has lived this problem. Scheduling a single investor–issuer meeting can take two to three days of emails across Sales, bankers and the issuer, while hundreds of meetings are still tracked by hand in Excel. Most of that coordination is routine, so we automated it, leaving bankers more time for clients.

Alex Xue: idea, workflow design and testing. Came up with Corda AI, mapped the conference scheduling process step by step, and built the core workbench: scheduling flow, banker coverage and meeting-room allocation, email import, and Excel/PDF exports. He also led testing against real scheduling scenarios. Tao Li: product and AI integration. Built the Nicrron model pipeline, the PRISM tracing, alerts and root-cause fix, the Ninoo deployment, and the demo video and pitch.

Corda AI runs on NicrronNinoo end to end. Every AI call goes through the Nicrron gateway (api.nicrron.ai), so one key reaches two models: deepseek/deepseek-flash sorts each incoming email quickly and cheaply, and anthropic/claude-sonnet-5 pulls out the meeting details (issuer, investor, time window, attendees). We imported the finished workbench into Ninoo Builder, ran it in the live preview, and published it with one click at https://corda-ai-ninoo-c084.ninoo.app.

PRISM watches the agent. Each email is a PRISM session with 8 traced tool calls: validate, dedupe, triage, match, extract and save. Alert rules track the ingest agent’s error rate and its failed runs. PRISM’s root-cause analysis grouped 17 failures from agent v1 into one cluster, which pointed us to the fixes: ignoring quoted email history, handling AM/PM correctly and matching short issuer names. Version 2 has no new failures on the same inputs.

The app is built with TypeScript, React, Next.js 15, Tailwind CSS, PostgreSQL and Prisma. Outlook .eml and .msg emails are imported with mailparser and msgreader, schedules export to Excel with ExcelJS and to PDF with PDFKit, and the demo video was made in Remotion.