Lodestar

Inspiration

Lodestar is an autonomous AI financial coach PoC for Shinhan Bank Vietnam's SOL app, built for Innoboost 2026.

The name reflects the product's role: a guiding star for customers navigating their financial lives — surfacing the right signals at the right time without ever telling them what to do.

Runs fully on-premise using Qwen3 via Ollama, with a proactive insight feed, cross-entity financial simulation across all four Shinhan entities (Bank, Finance, Securities, Life), a per-customer learning loop, and a Vietnamese-language first experience.

What it does

Not a chatbot. An autonomous agent that continuously monitors customer transactions, detects patterns and life events, surfaces ranked insight cards, and enables drill-down conversations with inline charts.

Key innovations:

Cross-entity financial digital twin — simulate "what if I buy a house?" across mortgage (Bank), existing loans (Finance), investments (Securities), and insurance (Life)
Predictive life event detection from transaction patterns (baby, home purchase, career change)
Per-customer learning loop with quality gating, semantic deduplication, and importance feedback
Federated cohort intelligence — anonymised patterns benefit similar customers (k-anonymity, differential privacy)
Van Tharp process/outcome separation — the system doesn't reinforce lucky advice
Deterministic MCP tools — all financial calculations done in Python, LLM only reasons and communicates
Workflow-as-tool orchestration — pre-built LangGraph workflows invoked by Qwen-Agent with ReAct fallback

How we built it

Backend: Python 3.11+, FastAPI, LangGraph, Qwen-Agent, Qdrant, SQLite, bge-m3 via sentence-transformers, underthesea (Vietnamese NLP), SDV + Faker (synthetic data)

LLM: Qwen3:14b via Ollama at localhost:11434 (production target: Qwen3-8B Q4_K_M via llama.cpp on 12GB GPU)

Frontend: Next.js 15 + shadcn/ui + Recharts, deployed on Vercel. Shinhan Blue (#0046FF) OKLCH theming.