Senior AI & ML Product Manager · FPT Software

Building AI products

Senior AI & ML Product Manager & Analytics Lead with 3–5 years specializing in AI/ML product development. Led 6 AI products across FPTShop and Nhà Thuốc Long Châu — from 0 to production. Reports directly to C-Level.

~3B
VND Revenue (Sep–Nov 2025)
435+
Daily AI Conversations
71%
Search Keyword Coverage
70%+
Ops Time Saved (SSC)
About Me
AI PM who bridges model capability
and business value.
Eddie Le

I'm not just a Product Manager — I bridge the gap between AI capability and business value. As AI & ML Product Manager and Analytics Lead at FPT Retail, I manage ROI & success metrics end-to-end and report directly to C-Level (CTO, CIO), Data Center Director & Head of AI.

I believe great AI products don't come from the best model — they come from understanding the right problem, measuring success correctly, and iterating fast based on real user data.

My work spans two dimensions: customer-facing AI that directly drives revenue and engagement, and internal AI that unlocks operational efficiency and intelligence at scale.

0-to-1 LaunchData-drivenLLM / RAGSemantic SearchAI AgentStakeholder MgmtGrowth & RetentionTeam Lead
Senior AI & ML Product Manager
FPT Software
2024 – Present
Led 6 AI/ML products from 0 to production. Responsible for Business Architecture and managing the full AI team. Manage ROI & Success Metrics. Report directly to C-Level (CTO, CIO), Data Center Director & Head of AI.
AI & ML Analytic Lead
FPT Software
2024 – Present
Directly managing: Business Analyst, Product Owner, Data Analyst, Data Scientist, Data Engineer, AI Engineer, AI Tester — end-to-end from research to deployment.
Product Architecture
6 AI products across 2 strategic pillars.
AI PRODUCT PORTFOLIO FPT Retail · 6 Products 🎯 Customer-Facing AI Revenue · Search · 3 products ⚙️ Internal Operations AI Efficiency · Intelligence · 3 products AI SALES ASSISTANT FPTShop ~3B VND · 435 conv/day 5 mo · LLM+RAG+Rules SEMANTIC SEARCH FPTShop 71% coverage · +28.7% CTR 6 mo · Vector+Fine-tune SEMANTIC SEARCH Long Châu (Pharma) ~70% coverage · Safety-first 12+ mo · Pharma NLP AI AGENT AUTOMATION SSC · FPT Retail 70%+ time saved · 50–200 users Agentic · NL Interface IQP — INTEL. QUALITY FPT Retail Internal 30–50% time saved 6–12 mo · KG+LLM AI SEO WRITER Marketing CMS ~200M VND/mo saved <3 mo · LLM+SERP SHARED TECH FOUNDATION LLM · RAG · Vector DB · Fine-tuning · Knowledge Graph · Vietnamese NLP · Rule-based Engine · SERP Crawling
Case Studies
Every product. Full story.
🎯
Customer-Facing AI
Products that directly touch end users — driving revenue, engagement, and search experience
3 Products
⭐ Featured · AI Sales Assistant
AI Sales Assistant
FPTShop Website · 5 months · 0 → Production
Problem
FPTShop had tens of thousands of SKUs but a low browse-to-purchase conversion rate. Customers lacked timely, personalized consultation. Human advisors were only available during business hours — leaving an enormous gap in the purchase journey.
Process
Month 1–2 · Discovery
User research, purchase journey analysis, intent gap mapping. Aligned stakeholders on success metrics across two distinct revenue streams.
Month 2–3 · Architecture
Designed hybrid approach: LLM + RAG + Vector DB + Fine-tuning + Rule-based conversation manager. Defined full pipeline: query → intent → response.
Month 3–5 · Build & Launch
Led DS/DE/AI Engineer team to build 2 revenue streams: AI direct orders (on-chat) + AI-to-human handoff pushing leads to sales advisors.
Key Challenges
  • Low data quality — required full catalog cleaning and normalization pipeline before model training
  • Initial model accuracy insufficient — iterated through 3 rounds of fine-tuning and evaluation
  • Stakeholder alignment on how to attribute and measure AI contribution within the sales funnel
Outcomes
~3B VND
Revenue Sep–Nov 2025 across 2 AI streams
435+
AI conversations per day (average)
43%
5-star conversation rate from users
Tech Stack
LLMRAGVector DBFine-tuningRule-based EngineConv. Manager
Key Learning
Hybrid architecture is decisive — no single approach is strong enough standalone. Rule-based ensures safety, LLM ensures flexibility, RAG ensures accuracy. Success comes from knowing when to use what.
Case Study · Semantic Search
AI Semantic Search
FPTShop Website · 6 months · 0 → Production
Problem
FPTShop's keyword-matching search required users to know the exact product name. Users searching by description ("good night camera phone under 10M") got zero results. With 153k unique search queries logged monthly, the gap was enormous.
Process
Month 1–2 · Discovery & Data
Analyzed 153k real user keywords from the log center. Classified intent: navigational, informational, transactional. Established zero-result and low-relevance baselines.
Month 2–4 · Model & Pipeline
Designed embedding pipeline for the entire catalog. Fine-tuned on Vietnamese e-commerce domain. Built hybrid retrieval: dense vector search + sparse keyword fallback.
Month 4–6 · Eval & Launch
Evaluated on 153k real keywords. A/B tested on live traffic. Refined ranking, re-ranking and query understanding before full launch.
Key Challenges
  • Product names with technical abbreviations — model needed domain-specific vocabulary
  • Users search by description ("long battery", "good selfie camera") — no exact keyword match
  • Vietnamese with/without diacritics — required robust query normalization layer
Outcomes
71%
Keyword coverage — from zero-result to relevant results across 153k real user keywords
+28.7%
Click-through rate increase after switching to semantic search
↑ Conversion
Search-to-purchase conversion improved post-launch
Tech Stack
Sentence EmbeddingsVector DBFine-tuningHybrid RetrievalRe-ranking
Key Learning
Eval set determines quality — 153k real keywords from user logs was the most valuable asset. Early investment in data annotation saves many iterations later.
Case Study · Pharma Search
AI Semantic Search
Nhà Thuốc Long Châu · 12+ months · 0 → Production
Problem
The pharmaceutical domain is significantly more complex: the same drug can have multiple brand names, active ingredients, and therapeutic indications. Users search by symptom, active ingredient, or misspelled names. A wrong recommendation has direct health consequences.
Process
Phase 1 · Pharma Knowledge Graph
Built pharmaceutical ontology: mapping active ingredient ↔ brand name ↔ symptom ↔ disease category.
Phase 2 · Domain Fine-tuning
Fine-tuned embedding model on Vietnamese pharmaceutical corpus with stricter eval than FPTShop.
Phase 3 · Safety & Accuracy Layer
Designed safety guardrails: no prescription drug suggestions for OTC queries, flagged sensitive queries for manual review.
Key Challenges
  • Drug names hard to spell correctly — "Augmentin", "Cefuroxime", "Amoxicillin"
  • Users search by symptoms, not knowing the exact drug name
  • Accuracy more critical than recall — wrong recommendation can harm users
Outcomes
~70%
Keyword coverage on Long Châu's real user keyword set
↑ Search UX
Users find medicine even with misspelled names or symptom-based queries
12+ mo
Longest timeline — reflects pharma complexity and high accuracy bar
Tech Stack
Pharma OntologyDomain Fine-tuningVector DBSafety GuardrailsKnowledge Graph
Key Learning
Safety-first is a product mindset, not a feature — pharma requires completely different recall vs precision tradeoffs. Every modeling decision has health consequences for real users.
⚙️
Internal Operations AI
Products that automate workflows, reduce manual effort, and enable data-driven decisions inside FPT Retail
3 Products
Case Study · AI Agent · Internal Tool
AI Agent Automation Workflow
SSC System · FPT Retail Internal Operations
Problem
FPT Retail's SSC system handled a large volume of internal workflows: order processing, periodic reporting, internal approvals. Most tasks were manual, repetitive, and consumed hours daily. Staff constantly context-switched between systems, prone to human error.
Process
Phase 1 · Workflow Mapping
Shadowed operations teams to map all manual workflows. Classified by frequency, time cost, error rate, automation feasibility. Prioritized highest-ROI workflows first.
Phase 2 · Agent Architecture
Designed AI Agent with natural language interface — users command in natural Vietnamese, agent executes multi-step workflows on SSC via tool-calling with internal APIs.
Phase 3 · Rollout & Adoption
Rolled out module by module, collected feedback from 50–200 internal users. Iterated rapidly — internal tools have very short feedback loops.
Key Challenges
  • Designing the agent robust enough to handle edge cases in production internal environment
  • Change management — shifting staff from established manual workflows to AI-assisted ones
  • Defining autonomy boundaries: AI self-decision vs requiring human approval for high-risk tasks
Outcomes
70%+
Reduction in manual processing time across automated workflows
50–200
Internal users actively using daily on the SSC system
Natural
Language interface — reports to operations in natural Vietnamese
Tech Stack
AI AgentTool CallingLLMNL InterfaceAPI IntegrationMulti-step Workflow
Key Learning
Internal tools are the best place to learn AI Agent design — short feedback loops, lower risk, users willing to iterate with the team. Lessons from SSC are being applied to future consumer-facing AI products.
Case Study · AI Platform · Internal Tool
IQP — Intelligence Quality Platform
FPT Retail Internal · 6–12 months · 0 → Production
Problem
QC/Testers spent approximately 2 story points per sprint just reading business documents and manually generating test cases — before any actual testing began. Testers frequently missed edge cases and had no structured way to clarify ambiguous requirements with BAs.
Process
Phase 1 · Discovery & Problem Framing
Shadowed QC/Tester workflows across teams. Mapped the full document-to-testcase pipeline. Identified the core gap: testers lacked a structured knowledge layer to reason about system complexity.
Phase 2 · Knowledge Graph Architecture
Designed an extraction pipeline that ingests business and technical documents, then builds a Knowledge Graph of nodes (features, entities, rules) and edges (relationships, dependencies).
Phase 3 · AI Agent + Test Generation
Built LLM-powered Q&A agent grounded on the Knowledge Graph. Added AI test case generation — producing raw test cases covering key business scenarios. Launched to 50–100 testers.
Key Challenges
  • Extracting accurate Knowledge Graphs from complex, inconsistently formatted business documents
  • LLM hallucination risk — required grounding guardrails and confidence scoring
  • Tester adoption: shifting from "read everything manually" to trusting AI-generated content
  • Uneven input document quality across teams — needed a doc quality signal before extraction
Outcomes
30–50%
Reduction in time reading docs & generating test cases — from ~2 story points to under 1
50–100
QC/Testers actively using IQP across FPT Retail teams
↓ Missed
Fewer missed business scenarios — AI test cases provide baseline coverage
↑ Clarity
Testers proactively clarify ambiguous requirements using AI-suggested questions
Tech Stack
Knowledge GraphLLM AgentRAGDoc ExtractionTest Case GenHallucination Guardrails
Key Learning
Knowledge Graph is the missing layer between raw docs and AI reasoning — plain RAG loses structural relationships. The graph forces AI to reason about dependencies, not just retrieve text chunks.
Case Study · AI Content · Marketing
AI SEO Content Writer
Marketing CMS · FPTShop · Under 3 months · 0 → Production
Problem
FPTShop's SEO content operation relied heavily on freelance writers — resulting in high production costs (~200M VND/month), inconsistent content quality, and limited scalability. The team had no systematic way to scale article output or maintain a coherent SEO strategy across hundreds of product categories.
Process
Phase 1 · Pipeline Architecture
Designed end-to-end LLM content generation pipeline: keyword input → SERP crawling → content structure extraction → AI generation → SEO optimization. Each stage was modular to allow rapid iteration.
Phase 2 · Prompt & Template Engineering
Built prompt frameworks and content templates ensuring SEO quality, brand voice consistency, and structural compliance (headings, meta, internal links). Enabled competitor analysis and keyword mapping as generation inputs.
Phase 3 · CMS Integration & Launch
Integrated the pipeline with FPTShop's internal CMS for seamless content creation and publishing workflows. Launched within 3 months — from 0 to fully operational at scale.
Key Challenges
  • Ensuring AI-generated content met Google's quality standards — required iterative prompt engineering and SEO validation layers
  • SERP crawling accuracy — extracting meaningful content structure from competitor pages for generation guidance
  • Stakeholder alignment on content quality threshold before reducing freelancer dependency
Outcomes
~200M
VND/month saved in content production costs — direct replacement of freelance writers
2×
Article output — from ~100 to ~200 articles/day with the same team size
<3 mo
Fastest 0-to-production timeline across all 5 products — shipped in under 3 months
↑ SEO
Improved control over SEO strategy and execution — reduced dependency on external writers
Tech Stack
LLM Pipeline SERP Crawling Prompt Engineering Content Templates CMS Integration Keyword Mapping SEO Optimization
Key Learning
Speed of execution matters more than perfection at launch — this was the fastest product shipped (under 3 months). The key was scoping tightly: solve one pain point first (cost reduction), prove ROI, then expand. Starting with a narrow prompt framework and iterating based on real SEO performance was far more effective than designing the "perfect" system upfront.
AI Product Thinking
How I approach AI products.
📐
Defining AI success metrics
AI can't be measured in story points. I define success across 3 layers: model metrics (accuracy, latency, hallucination rate), product metrics (engagement, satisfaction, conversion), and business metrics (revenue, cost saved). All three must align.
🔬
Working with Data Scientists
Start with problem framing, not model selection. I help DS understand business context, establish eval frameworks before training, and translate model outputs into user-facing decisions. PM must be technical enough to review the pipeline.
⚖️
Responsible AI in practice
For AI Sales Assistant, I designed guardrails for hallucination and human-in-the-loop when confidence is low. For Long Châu, safety-first meant strict recall vs precision tradeoffs. Responsible AI is a product requirement — not optional.
🔄
Hybrid Architecture Thinking
5 products taught me: there is no silver bullet. LLMs excel at language but can't be trusted for facts. RAG solves grounding but requires strong data pipelines. Rule-based ensures predictability. Good PMs know when to combine — not follow hype.
Skills & Tools
What I bring to the table.
AI / ML Product
LLM IntegrationRAGVector SearchFine-tuningPrompt EngineeringAI Agent DesignSemantic SearchNLP
Product Management
0-to-1 LaunchRoadmappingPRD WritingOKR / MetricsA/B TestingUser ResearchRICE / MoSCoWStakeholder Mgmt
Data & Analytics
SQLMixpanelFunnel AnalysisRetention AnalysisCohort AnalysisModel Evaluation
Team & Leadership
AI Team LeadBusiness ArchitectureCross-functionalAgile / ScrumOKR Cascading
Design & Tooling
FigmaJiraNotionConfluenceMiroPostman
Domain Expertise
E-commerceRetail TechPharma / HealthConversational AISearch & DiscoveryWorkflow Automation
Contact
Let's build
something great.
Open to Senior AI PM / AI Product Lead opportunities. Happy to connect and chat.
My Email Visit My LinkedIn Welcome To My Blog