The Next Generation of Wealth Advisory

The guiding star for the wealth manager.

---

Polaris is a wealth-manager (WM) workbench for the SIX · Noumena · NTT Data challenge "The Next Generation of Wealth Advisory." It turns the firehose of market news into a focused, per-client action queue: the WM sees what changed for each client, understands why (news → analysis → advice → action), and acts in one click — while the client always decides.

Three apps work together as one product:

Folder	What it is
frontend/	React + Vite + TypeScript SPA — the WM dashboard
server/	Express + SQLite backend — the UML entities + WM use cases, with real persistence
system/	LangGraph multi-agent system — the inference that produces news, analysis and suggestions

The full loop is live and integrated: the agent graph reasons over each client's DNA, the CIO mandate and live news, writes suggestions into the workbench, and the WM applies or rejects them from the dashboard — every decision flowing back to evaluate plan quality.

---

Architecture

        data/*.xlsx ──seed──▶  SQLite (server/data/workbench.db)
                                     │
   ┌───────────────────────┐        │  REST  /api/*
   │  system/ (LangGraph)   │◀──────▶│
   │  multi-agent inference │   Express API (server, :4000)
   │  • client DNA          │        │
   │  • CIO plan            │        │  /api/*  (Vite proxy)
   │  • news (GraphRAG)     │        ▼
   │  • create / explain    │   React SPA (frontend, :5173)
   │  • evaluate plan       │        │
   └───────────────────────┘        ▼
                                 Wealth Manager
                              (applies / rejects)

• Frontend → API: frontend/src/lib/api.ts (fetch) + lib/queries.ts (react-query). Mutations invalidate dashboard/customer queries so the UI refetches automatically.
• API → DB: server/src/index.ts (routes), server/src/repo.ts (drift + apply/drop), server/src/schema.ts (DDL), server/src/seed.ts (xlsx → rows).
• System → API: the LangGraph agents read client/portfolio/CIO data from the workbench API, retrieve news through GraphRAG, and write back suggestions and tailored client messages. See system/README.md for the agent graph.

---

Run it

Prerequisites

• Node 22.12+ and npm (.nvmrc pins 22.12.0 in frontend/)
• Python 3.11+ (for the advisory system)
• Docker (for GraphRAG)

1. Backend — seeds the SQLite DB from data/*.xlsx on first boot

cd server
npm install
npm run dev            # http://localhost:4000

2. Frontend — Vite dev server, proxies /api → http://localhost:4000

cd frontend
npm install
npm run dev            # http://localhost:5173

Open http://localhost:5173.

3. Advisory system — the multi-agent inference

cd system
cp .env.example .env   # set your LLM endpoint + key
pip install -e ".[dev]"
docker compose up -d   # GraphRAG
python -m src.main

To rebuild the database from the workbooks at any time:

cd server && npm run seed -- --reset

A single-origin production image (frontend served by the backend) is provided via the root Dockerfile and render.yaml.

---

The use cases (UML)

Use case	Endpoint
Get dashboard data (+ TLDR news)	GET /api/dashboard
Get client	GET /api/customers/:id
Add client data (comment)	POST /api/customers/:id/comments
Apply suggestion (records a transaction, swaps the holding)	POST /api/suggestions/:id/apply
Drop suggestion	POST /api/suggestions/:id/drop
Run advisory pipeline for a client	POST /api/system/trigger
Submit WM decision (HITL)	POST /api/system/hitl

Per the brief, news is shown on the WM dashboard, not inside the client view (even though the UML lists "get TLDR news" under "get client").

---

App map

• Dashboard (/) — clients overview, action queue (pending suggestions), market signals (TLDR news). frontend/src/pages/DashboardPage.tsx
• Clients (/clients) — the book. ClientsPage.tsx
• Client detail (/clients/:id) — tabs: Holdings (donut + table), Suggestions (apply/drop), History, Relationship · DNA (CRM log + add data). ClientDetailPage.tsx
• Action queue (/alerts) — cross-book pending suggestions.

The advisory graph (system/) drives eight agents end-to-end: get_client_dna → get_cio_plan → get_current_plan → fetch_news → check_actions → create_plan → explain_plan → evaluate_plan, with the WM in the loop between explain and evaluate.

---

Design

Palette follows the SIX reference UI: light neutral canvas, white cards, SIX red as the accent (active nav, primary actions, alerts), green for positive figures. Tokens live in frontend/src/index.css. Type: Bricolage Grotesque (display) · Inter (body) · IBM Plex Mono (tabular figures). Brand assets live in frontend/public/brand/.

Polaris — the WM's true north.