Open to new opportunities
Songyu Bao
Senior AI Education Product Manager
I rebuild fragmented, complex businesses into scalable products — and use AI to turn messy human experience into structured, explainable decisions.
0
Years Experience
0
Revenue Supported by Systems I Led
0
Growth Achieved
0
Faster AI-Driven Delivery
Scroll to explore
About Me
Complex System Architect & AI Efficiency Advocate
I'm a Senior Product Manager with 8+ years of experience in the EdTech industry, currently at NetEase Youdao . I specialize in complex system decoupling & architecture design , turning fragmented legacy systems into unified, scalable platforms that drive billions in revenue.
My unique edge: I bridge traditional product engineering with AI-driven efficiency tools . I've compressed product delivery cycles from 1-2 weeks down to 2-3 days using custom AI workflows built on Claude Code, Codex, and Kimi Code. I'm pushing toward becoming a full-stack product manager who can design, prototype, and ship independently.
Beyond building products, I mentor team members, transforming personal AI workflows into shared organizational infrastructure.
4
Major Systems Built
17
Core Modules Designed
5x
AI Efficiency Gain
0
Critical Incidents in Migration
Product Philosophy
How I Think About Products
The principles that guide every architectural decision and product strategy I make.
AI is a Leverage, Not a Replacement
AI doesn't replace human judgment — it amplifies it. I use AI to compress delivery cycles and free up time for deeper thinking on product strategy.
Proof: 1-2 weeks → 2-3 days delivery cycle
Individual Capability → Team Infrastructure
A personal AI workflow only multiplies one person. The real value is turning it into shared infrastructure that levels up the entire team.
Proof: 3 team bots deployed across department
Data-Driven, Not Intuition-Driven
Every product decision should be validated by data. I build measurement frameworks first, then iterate based on real user behavior and business metrics.
Proof: ~1,000 class transcripts calibrated via Codex into a scoring model
Career Journey
From Design to Product Leadership
A trajectory of continuous growth in China's EdTech industry, with one thread throughout: building AI education products since 2018.
2021.11 - Present
NetEase Youdao
Senior Product Manager
Lead PM for the 1-on-1 business systems — drove a full architecture rebuild, shipped AI diagnosis products, and now leads AI Agent & PM Spec Engineering work.
System Architecture
AI Agent
PM Spec Engineering
Team Lead
Midplatform
Revenue: multi-billion ¥ scale
·
Strong YoY Growth
·
Rebuilt 3 legacy systems into 1 platform
·
Shipped 3 AI diagnosis products
·
24h Learning Assistant Agent (Demo)
·
PM Spec Engineering G0–G6 in pilot
·
AI workbench cut delivery to days
·
Zero-incident high-risk migration
·
Star Project Award 2022
2021.02 - 2021.11
Zuoyebang Education
Senior Product Manager
Owned acquisition-and-conversion products and the MIS back office — rebuilt SCRM and WeCom tooling, and restructured the back-office architecture.
User Acquisition
SCRM Platform
WeCom Ecosystem
Drove 30%+ incremental new user growth
·
Built full WeCom automation funnel
·
Visual user profiles in the sales workflow
·
Public-to-private domain conversion
·
Restructured MIS back-office architecture
2019.08 - 2021.02
Doushen Education
Senior Product Manager
Led part of the product line with a small team — built the dual-teacher SaaS 0-to-1, standardized academic operations, and rebuilt the flagship app's order flow.
Dual-Teacher SaaS
Admin System
Mobile App
Supported 10K+ teachers on SaaS platform
·
3x improvement in operation efficiency
·
Built dual-teacher SaaS 0-to-1 across 3 ends
·
Standardized scattered ops into one back office
·
Rebuilt flagship app's full order flow
2018.06 - 2019.08
iFLYTEK
Interaction Designer (Product-oriented)
Where my AI education track record began — drove an NLP/OCR grading assistant and a speech-recognition assessment app from 0 to 1.
NLP/OCR Products
AI Evaluation
User Research
Participated in 2 AI product launches
·
Improved user satisfaction by 18%
·
Shipped NLP/OCR grading for a gov client
·
Designed ability model behind AI scores
AI + Education
My Core Insights
A few convictions from years of building AI products in education — each one earned from the projects shown below, not from theory.
AI Amplifies Teachers, Doesn't Replace Them
The best AI products in education don't aim to replace teachers. They free teachers from repetitive work (grading, paperwork) so they can focus on what humans do best — inspiration and mentorship.
Personalization Needs Constraints, Not Endless Options
Too much personal choice leads to decision paralysis. Good AI personalization gives learners the right next step at the right time, not 10 options to choose from.
Transparency Beats Black Box
In education, people need to understand why AI made a recommendation or score. A slightly less accurate but fully transparent model always wins out in production over a black-box.
Selected Work
The Work Behind the Ideas Above
The views above only matter if they hold up in real products. My work runs on two threads: rebuilding complex systems at scale, and using AI to turn unstructured human experience — an exam, a sales call, a recorded class — into structured, explainable, actionable diagnosis. Filter by type below.
2024 - 2025
1-on-1 Business System Rebuild
Rebuilt three fragmented legacy systems into one modular platform carrying multi-billion ¥ annual revenue.
Problem: Three legacy systems, three definitions of user/order/fulfillment — seamless outside, fractured inside.
Action: Chose a full rebuild around one unified domain model; layered the system and decoupled RBAC/payment/identity; designed a staged migration over a single cutover to protect live revenue.
Result: One modular platform supporting multi-billion ¥ revenue, faster delivery, zero critical incidents during migration.
Key Takeaway
Designing large-scale architecture from first principles, and de-risking a migration on a system that can't go down.
¥B-scale
Annual Revenue
High
YoY Growth
60+
Sub-modules
2023 - 2025
B/C Integrated Learning Platform
Built the core B2C platform from the ground up, carrying the majority of company revenue.
Problem: B-side and C-side ran on disconnected systems; the journey broke at every handoff, with no unified data to personalize or to see revenue leaks.
Action: Owned it from 0 to 1 — designed a full-stack architecture connecting scheduling → learning → payment into one loop, with an event-tracking layer for data-driven decisions.
Result: Became the core revenue platform with stable multi-year growth; the unified data loop made personalization and revenue analysis routine. Recognized as a company Star Project.
Key Takeaway
Owning a revenue-critical product end-to-end from zero, and building the data foundation that makes a platform measurable.
Core
Revenue Platform
0→1
Built from scratch
Star
Project Award
2023
New Exam-Prep Business · Multi-end Product
Designed the full product chain for a new exam-prep business across multiple endpoints — a new ¥100M-scale growth curve.
Problem: A new model spanning many endpoints with no unified data layer — inconsistent experience, no clear signal on what worked.
Action: Led full-chain product design across all endpoints; redesigned the mid-platform data structure for one source of truth; aligned three teams onto a single roadmap.
Result: Launched coherently across endpoints at ¥100M-scale estimated revenue, opening a new growth curve — and a repeatable playbook for standing up a new business line.
Key Takeaway
Taking a 0-to-1 multi-endpoint business to coherence by fixing the data layer first and aligning teams.
¥100M-scale
Est. Revenue
2024
Part-time Creator Distribution Platform
A platform turning part-timers' own social accounts into a distributed acquisition channel — task dispatch in, leads out, settled automatically.
Problem: Rising acquisition cost; wanted to use part-timers' social accounts as a low-cost channel, but had no way to dispatch, verify, and settle at scale without manual disputes.
Action: Designed around two hard parts — task dispatch (bind accounts, route posting & lead tasks to the right people) and automatic settlement (define how each task is verified and priced, tying payout to verified results with a clear audit trail).
Result: A repeatable, low-cost, self-running acquisition channel — self-serve part-timers, auto dispatch & settlement, transparent cost-per-lead.
Key Takeaway
Turning a fuzzy growth idea into a closed-loop system, nailing the two hardest links: fair dispatch and result-tied settlement.
Live
Growth Channel
Auto-settled
2024
Product Exchange Capability
Built exchange logic for course packages and group classes so the business can flexibly swap purchased products.
Problem: Paid course packages/classes often need post-purchase changes; without an exchange flow, every change meant manual refund-and-repurchase that broke orders and fulfillment.
Action: Designed a dedicated exchange flow instead of refund+repurchase — maps original value to the new product, keeps the order continuous, and keeps revenue and fulfillment in sync.
Result: A flexible self-serve exchange that keeps orders, revenue and fulfillment consistent, sharply cutting manual workarounds.
Key Takeaway
Seeing a fuzzy ask as a transaction-integrity problem and solving it at the order layer, not patching the symptom.
2024
App Home Revamp for Acquisition
Rebuilt the app home into a configurable acquisition surface for multi-channel app-store campaigns.
Problem: Paid installs came from many app stores, but the static home gave every channel the same generic screen with no guided next step, wasting traffic.
Action: Turned the home into a configurable surface — banner slots, an icon-grid zone and targeted pop-ups swappable without a release — so marketing can match content to each channel independent of the ship cycle.
Result: Multi-channel traffic lands on tailored content with a clear next step; campaigns launch without engineering, and the home became a reusable conversion lever.
Key Takeaway
Designing the home as configurable infrastructure for the people who run campaigns, so growth isn't bottlenecked on engineering.
2024
In-WeChat Mini-Program Acquisition Funnel
Built a mini-program funnel for in-WeChat campaigns, shortening the path and lifting conversion.
Problem: In-WeChat ads pushed users out to H5 or an app download before converting; every extra hop leaked users, so in-WeChat traffic converted poorly.
Action: Built a mini-program funnel that keeps the whole journey inside WeChat — ad → mini-program → conversion, no ecosystem exit or install — and designed the entry links to minimize steps from click to conversion.
Result: A materially shorter path and higher in-WeChat conversion, making official-account traffic a far more efficient acquisition source.
Key Takeaway
Reading conversion as a path-length problem and removing hops where users actually are, not dragging them to your platform.
2024
API Auto Order-Sync Pipeline
Connected third-party commerce channels via API so orders sync back in real time — no manual entry.
Problem: Third-party channel orders lived outside our system and had to be manually re-entered — slow, error-prone, unscalable as channels multiplied.
Action: Designed an API pipeline integrating multiple third-party platforms with real-time order push; the hard part was reconciling each platform's order model into one consistent internal format.
Result: Orders from every channel now flow in automatically and in real time, removing manual entry; adding a channel became configuration, not new manual work.
Key Takeaway
Designing an integration layer that absorbs many external platforms into one clean internal model, so channel growth doesn't create operational debt.
讯飞 · 2018-2019
AI Essay-Grading Assistant (NLP/OCR)
Built a 0-to-1 AI grading tool on NLP/OCR for a government education client — my entry into AI education, in 2018.
Problem: Manual essay grading is slow, inconsistent, unscalable; a government client needed automated assessment people would trust.
Action: Drove the tool from 0 to 1 on NLP/OCR — designed the grading workflow and interaction, aligned closely with the G-side client on what 'trustworthy' meant, and saw it through to launch.
Result: Shipped and turned manual grading into a scalable, consistent process — and where I first learned that AI usefulness depends on trust, not just accuracy.
Key Takeaway
My AI-education track record starts in 2018, not with the recent wave — and with the hardest customer to win: a government client.
讯飞 · 2018-2019
Spoken-Language Assessment App
An AI app scoring spoken language via speech recognition — designed the evaluation flow, feedback and the ability model behind the score.
Problem: Spoken-language practice is hard to assess at scale; a raw score tells learners nothing about what to fix.
Action: Designed the evaluation flow on speech recognition and decomposed the score into an ability model (pronunciation, fluency…) pointing to a concrete next step, iterated through user research and usability testing.
Result: Learners got specific, actionable feedback rather than an opaque number — reinforcing that explainable beats merely accurate.
Key Takeaway
Turning a raw AI score into an actionable ability model — the 'explainable, actionable' instinct I apply to every AI product today.
豆神 · 2019-2021
Dual-Teacher Classroom SaaS
Built a 0-to-1 dual-teacher platform across web, pad and mobile — enrollment, live interaction and homework in one system.
Problem: The dual-teacher model spans many roles and three device types; without one platform, the classroom experience was fragmented and unscalable.
Action: Researched the full flow, set product goals and modules, and scoped a phase-one launch around the highest-value loop (enrollment → live → homework) across all three ends.
Result: The classroom ran on one coherent three-end SaaS platform supporting a large teacher base, turning a fragmented operation into a repeatable one.
Key Takeaway
Scoping a sprawling multi-role, multi-device system into one shippable phase-one loop — cutting well matters as much as building.
豆神 · 2019-2021
Academic Operations Backend
Turned scattered academic operations into standardized modules integrated with company permission and CRM systems.
Problem: Academic operations ran on ad-hoc habits, not shared standards, so nothing could integrate with company permission or CRM systems.
Action: Researched how teams actually worked, distilled it into standardized SOPs, turned those into operations modules, and integrated them with company permission and CRM systems.
Result: Fragmented operations became one standardized, integrated backend giving consistent, scalable support and far smoother cross-team work.
Key Takeaway
Reading messy workflows, distilling them into standards, and encoding those into a shared system — the core of back-office product work.
豆神 · 2019-2021
Flagship Learning App
The company's flagship parent-facing app — owned enrollment, renewals, recommendation and the full order flow.
Problem: The flagship app didn't match how parents actually decide, and the selection-to-payment flow had gaps that leaked conversions.
Action: Restructured enrollment, renewal and recommendation around the course system and parents' real habits, and closed the full order loop from selection and payment to back-end processing.
Result: A coherent parent journey with a complete, reliable order flow — supporting core enrollment and renewal as the flagship consumer touchpoint.
Key Takeaway
Owning a flagship consumer app where every gap costs revenue — designing around real behavior, not internal structure.
作业帮 · 2021
WeCom Ecosystem Sales Tooling
Built WeCom sales tooling that surfaces a visual user profile beside every conversation.
Problem: Sales were converting inside WeCom but talking blind — no view of who each person was; generic outreach converted poorly.
Action: Used the WeCom open API plus key behaviors to render a visual user profile in the side panel, with a tag system, giving sales an in-flow tool and laying groundwork for automated segment-based outreach.
Result: Sales could see who they were talking to and tailor outreach, lifting effectiveness and turning scattered outreach into a systematic motion.
Key Takeaway
Putting the right data in front of the right person at the moment of action — a tooling instinct that later shaped my AI sales-tiering work.
2026
24h Learning Assistant Agent「Xiaoling」
AI covers high-frequency, rule-based student questions so human planners focus on high-value service — plus 24h learning companionship for paying users
Problem: High-frequency student questions relied on human planners; rule-based, low-value issues consumed staffing, and nights / off-hours went unanswered. Goal: let an AI Agent take the rule-based share, free planners for high-value service, and give paying users 24h learning companionship.
Action: Mined 100K evenly-sampled first-turn sessions and identified 71,470 issues (68,073 classified) by rule/non-rule attribute and time slot; built a night-coverage capacity model (peak 37–43 planners, down to 24–28 if non-rule transfer drops 50%→30%); designed 7 scenario types with knowledge/fact/judgment boundaries, NPS prompts and planner handoff.
Result: Interactive demo completed; “what it can answer, when to query data, when to hand off to a human” is now a reviewable rule set instead of guesswork.
Key Takeaway
Quantifying where AI should act from real service data before building the Agent — data-driven capability boundary and capacity modeling.
71,470
Issues Identified
7
Scenario Types
In Development / Demo
AI + Human Handoff
2026
Parent AI Live Digital-Human Agent
Upgrade one-way learning reports into a trustworthy, correctable, actionable conversation for parents
Problem: Traditional learning reports are one-way displays — parents can read but not ask, and still don't know what to do next. Goal: turn “report display” into a trustworthy, correctable, actionable parent-communication Agent.
Action: Designed a 13-turn semi-structured dialogue with digital-human state videos and explicit AI identity; scoped the V6.0 demo into an APP V1.0 with its own G0–G6 spec gates, source registry and demo real/mock boundaries.
Result: Multi-round interactive demo polished on mobile, forming a full journey from sales touch → H5 → account → Agent conversation; currently in spec phase.
Key Takeaway
Designing conversational trust for parents — companionship with explicit AI identity, correctable content, and a disciplined demo-to-app engineering path.
13
Dialogue Turns
Voice + Text
Input Modes
Personal Project
V6.0 Demo
APP Spec
2026
PM Spec Engineering
A business-agnostic spec-engineering method: every requirement keeps a single source of truth from research to acceptance
Problem: With AI in product work, generation is fast — but fact sources are unclear, requirements drift, demos diverge from PRDs, and acceptance boundaries stay fuzzy. Goal: a business-agnostic spec-engineering method where every requirement is traceable, rewritable and verifiable.
Action: Built a system-level spec root with G0–G6 gates — source registry, Evidence Ledger, Decision Log, workflow-state and change write-back; Claude Code / Codex / OpenSpec turn business material into proposals, specs, tasks and check scripts, separating structural checks, demo review, automated evidence and human acceptance.
Result: The auto-invoicing pilot passed G0/G1 and is advancing to G2; OpenSpec strict and custom checks pass; gate semantics and project-entry rules are now the shared skeleton for follow-up projects.
Key Takeaway
Treating AI-assisted product work as engineering — traceable, rewritable, verifiable and handoff-ready.
G0–G6
Stage Gates
G2
Pilot Progress
Method Engineering
Reusable Template
2024 - 2026
PM AI Workbench · Infinite Canvas
An infinite-canvas workbench where research, evidence, specs and review move forward on one canvas
Problem: PRD, reviews, competitive analysis and data insight were scattered across tools and personal experience — context broke between steps, and delivery speed & quality depended on individual state.
Action: Built an infinite-canvas PM workbench that organizes research, evidence, specs and review on one shared canvas; packaged PRD Review Bot, Competitive Analysis Bot and Data Insight Bot to support fast requirement-to-demo validation.
Result: Selected requirement & demo validation cycles compressed from 1-2 weeks to 2-3 days; team members now reuse one review and output standard.
Key Takeaway
Canvas-based context organization, AI toolchain integration, and turning personal practice into team infrastructure.
5x
Faster Delivery
3
Team Bots Built
2025 - 2026
AI-Powered Exam Paper Analysis
Students upload papers & answer sheets; AI produces a multi-dimensional personalized report
Problem: Traditional paper review relies on manual, question-by-question analysis — coarse, slow, impossible to scale to per-student personalization.
Action: Designed an analysis framework ending in a student's next-step decision; AI decomposes the paper across error-cause, error distribution, must-win questions, and time-allocation, aggregated into one actionable report.
Key Takeaway
Productized teachers' grading expertise into a scalable AI diagnosis, designed so users can act on it.
How it works
Input
Exam paper + answer sheet
AI Decompose
Error cause · distribution · must-win · time
Output
Actionable score-up report
Live
C-side Education
Explainable
2025 - 2026
AI Sales Tiering & Growth Model
AI analyzes sales talk patterns + real performance to define tiers and recommend strategies
Problem: Sales evaluation relied on subjective judgment; great talk tracks were never systematized.
Action: AI extracted linguistic & behavioral features from real conversations, cross-referenced with performance to derive tier profiles; landed in a growth system and a talk-strategy recommendation layer.
Key Takeaway
Made top reps' tacit experience explicit and replicable through behavior-to-outcome attribution.
How it works
Input
Sales talk tracks + real performance
AI Attribution
Cross behavior features with outcomes
Output
Tier model + talk recommendations
Live
Internal / B-side
Reusable
2025 - 2026
AI Teacher Performance Analysis
From ~1,000 class transcripts, AI turns black-box scoring into reviewable metrics & evidence clips for teacher coaching
Problem: High-quality teaching know-how stayed locked in class recordings — manual review doesn't scale, and its conclusions can't be re-checked or reused.
Action: Extracted teacher/student behavior features from ~1,000 real class transcripts; turned black-box scoring into business-reviewable metrics and evidence clips, producing actionable flow guidance and talk-track coaching.
Key Takeaway
Turned tacit teaching craft into reviewable, replicable improvement suggestions — explainable-AI design that a business team can actually read and adjust.
How it works
Input
~1,000 class transcripts
AI Analysis
Behavior features → reviewable metrics & evidence
Output
Flow guidance + talk-track coaching
Live
Explainable
Business Adopted
2026
PM Workstation · Idea to Live
A PM-native full-stack demo pipeline: from requirement to online deployment
Problem: PMs usually stop at PRD/prototype and depend on engineering for delivery, causing validation lag.
Action: Built a PM Workstation flow: AI-assisted requirement draft → prototype generation → vibe coding implementation → cloud deployment and feedback loop.
Result: Turned concept validation from 'cross-team waiting' to 'self-driven shipping', significantly shortening pilot cycles.
Key Takeaway
PM engineering literacy, AI-tool orchestration, fast end-to-end demo delivery.
Live
Deployment Status
E2E
From PRD to Prod
My AI Stack
How I Supercharge Productivity
From requirement to deployment, AI is embedded in every step of my workflow. Here's my actual stack.
01
Requirement & Research
Claude
Kimi
豆包
Use Claude, Kimi and Doubao to digest competitor docs, user feedback and market reports — synthesized insights and gap analysis in minutes; pick the tool per task instead of binding to one platform.
02
PRD & Spec
Claude Code
Codex
OpenSpec
Draft PRDs with clear user stories and acceptance criteria in Claude Code and Codex; OpenSpec keeps proposal, design, specs and tasks consistent and verifiable.
03
Build & Ship
Claude Code
Codex
Kimi Code
Claude Code, Codex and Kimi Code for full-stack implementation, API integration and deployment — turning a spec into a running product.
04
Analyze & Iterate
Claude Code
Codex
WorkBuddy
Feed metrics and feedback back in, then iterate directly in Claude Code, Codex and WorkBuddy — analysis and the next build on one loop.
Team Infrastructure
Beyond personal use of AI tools, I package repeatable capabilities into team-level infrastructure.
Shared AI Bots
Built PRD Review Bot, Competitive Analysis Bot and Data Insight Bot for PM team daily usage.
Tooling LayerWorkflow Standardization
Aligned requirement intake, PRD output and review checklist so every PM can run the same high-quality process.
Process LayerKnowledge Reuse
Converted scattered cases into reusable prompt/template library, reducing repeated communication and onboarding costs.
Knowledge LayerTools & Methods
What I Depend On
The frameworks and tools that shape how I think about product problems and drive execution.
需求分析
用户旅程地图
RICE 优先级
KANO 模型
影响地图
产品规划
北极星指标
OKR
分层 roadmap
用户价值-商业价值矩阵
数据分析
同期群分析
A/B 测试
漏斗分析
相关性分析
AI 工具
Claude Code / Codex
Kimi Code / Kimi Work
WorkBuddy / 豆包
OpenSpec
协作落地
用户故事
冲刺规划
影响地图
回顾总结
架构设计
领域驱动设计
依赖倒置
上下文映射
API First
PM AI Lab
Method Validation & Rapid Prototyping
Unlike the business-impact projects above, this section focuses on my independent lab: validating PM+AI methodologies, building runnable demos, and stress-testing toolchains before scaling them into team practice.
2025 - Present
Workflow Experiment
PM Workstation Demo Suite
Built a PM-native delivery sandbox where one person can complete requirement definition, interaction design, coding, and release. The goal is to validate the boundary of 'Product Manager as Builder' under AI collaboration.
Codex
Claude Code
Vibe Coding
Cloud Deploy
2025
Template Productization
Writing Workbench (AI PRD Studio)
Created a PRD-focused writing workspace for PMs: scenario prompts, requirement skeletons, review checklist, and quality gates in one flow. Used to reduce blank-page friction and improve first-draft quality.
PromptOps
RAG
PRD
Copilot Flow
2024 - Present
Toolchain Practice
Heavy-Use AI Toolchain Practice
Systematic daily use of Claude Code, Codex, Kimi Code and other copilots across requirement analysis, code generation, QA and deployment — tool picked per task, not bound to one platform. Continuous comparison and tuning to form reusable best practices.
Claude Code
Codex
Kimi Code
Workflow Ops
Expertise
Skills & Capabilities
A product manager who bridges engineering depth with AI innovation.
Product & Engineering
Domain Knowledge
AI & Efficiency
Tools & Software
Axure
Figma
Sketch
Claude
Claude Code
Codex
Kimi Code
WorkBuddy
Feishu/Lark
JIRA
Confluence
SQL
Python
HTML/CSS
Tower
Let's Connect
Looking for an AI-Savvy PM ?
Whether it's product strategy, AI integration, or EdTech innovation — I'd love to chat. Let's explore how I can bring value to your team.
Email
18501003630@163.com
WeChat
18501003630
Location
Beijing, China