---
title: "Unique Friend | Moxiang Tech: From Expert Systems to BERT — Banking Vertical LMs"
author: "Unique Research"
sourcePublication: "Unique Research (FFCAP English library)"
originalPublishedAt: "2023-05-05T19:11:27+08:00"
canonical: "https://ffcap.cn/en/research/20230505a085ua00"
source: "https://ffcap.cn/en/research/20230505a085ua00"
language: "en"
cnSourceUrl: "https://view.inews.qq.com/a/20230505A085UA00"
---

# Unique Friend | Moxiang Tech: From Expert Systems to BERT — Banking Vertical LMs

_Unique Research · Insight / Unique Friend profile · May 5, 2023 · Guest company: Moxiang Tech (摸象科技)_

Inventory type remains **insight** (Unique Friend / 非凡伙伴 series — not partner welcome). Seeking Series B; Unique Capital as financial advisor. Originally via 36Kr.

## Context

AI 1.0 NLP wave → translation/CS/marketing. Today LLM breakthrough reframes survivors: not “find a scene” but “do we rebuild with LLMs in the scene we already own?” Peers (e.g. Zhuiyi, Emotibot) already moved to BERT post-2018; ChatGPT wave is the next test.

**Moxiang** founded **2007**; from 2016 pivoted big-data precision marketing → “AI+finance” digital employees. Product: conversational **Xiangxiaomi** AI-employee platform + model-training for bank marketing/ops — SaaS annual or private deploy.

## Tech

2020: introduced Google BERT atop prior deep-learning + domain KB — fine-tune on finance data, compress structure, raise concurrency → banking vertical models; per-bank / per-scene proprietary models with regular refresh. CEO **Gao Peng**: >**150** pretrained bank-business vertical models; billions of real dialog turns/year. Why BERT: classic KB bots good on specialist Qs but weak on general intent/common-sense — hard to fix with KG alone.

## Product & economics

Xiangxiaomi integrates into bank workflows — web / WeCom / phone. Helps RMs filter cohorts, proactively pitch fitting products, hand off warm leads; auto-logs RM–customer talks into memos/reports. After 2–3 days scene training: NLP recognition >**95%**; ROI vs pure human marketing **5–10×**. Pricing: per AI-employee seat / year; exploring outcome bonuses. Clients: CCB, ICBC, BOC, PSBC, Industrial, SPDB + dozens SOE/joint-stock banks. Profitable; repurchase **90–95%**.

**TAM:** ~6 majors × ~40 provincial branches + 12 joint-stocks × ~40 + ~1,500 city/rural banks ≈ **~2,200** addressable entities. Goal by 2025: +100 banks/yr and **20,000** AI employees (now ~8,000 at branches; >10,000 by end-Q2’23).

Private deploy mandatory under finance data rules — limits some proprietary data for continuous vertical iteration; per-client private model ops raise cost. Gao: banks still want desensitized corpus training; Moxiang built one-click upgrade so bank IT can self-ops. Evaluating open LMs for private deploy vs closed general models.

## Team

Gao Peng PhD — student of ZJU academician Pan Yunhe; creator of China’s first mobile BOSS billing model; director/chief scientist of ZJU–Moxiang Data Intelligence Joint Lab (with ZJU AI Institute + data analytics center, 2020). Core from ZJU AI, Alibaba, ZTE. **17** AI patents; **78** software copyrights.

_1 CN image; attach on ingest._

— END —

---

Editorial English edition published on UniqueCapital: https://ffcap.cn/en/research/20230505a085ua00
On-site reading page: https://ffcap.cn/en/research/20230505a085ua00