VETC AI Agent Platform

The AI Copilot for Every Vietnamese Car Owner. A multi-agent AI system that unifies every car service into a single conversational interface. One message. Six expert agents working together.

Comment

VETC AI Agent Platform

Car Owner Copilot — Pitch Deck

Slide 1 — Headline

The AI Copilot for Every Vietnamese Car Owner

Turn the mandatory VETC e-tag into the smartest assistant a driver has ever had.

Slide 2 — The Problem

Vietnam has 4.5M+ registered cars and growing. Every car owner deals with the same fragmented chaos:

Pain Point Reality Today
Toll & trip costs Manually check VETC app. No context, no history.
Car wash WhatsApp the garage. Hope for a slot.
VETC wallet Runs out mid-highway. No warning.
Traffic fines (phạt nguội) Discovered by accident — weeks late, accruing interest.
Warning lights Google "đèn cá vàng là gì". Get Reddit.
Maintenance Call 3 garages to find one with availability.

The car ownership experience in Vietnam is broken into six disconnected apps, two hotlines, and a prayer.

Slide 3 — The Insight

Tasco already owns the fix. They just haven't connected the dots.

Tasco Group owns:
  ✓ VETC e-tag  →  real-time GPS + toll + mileage data on 2M+ vehicles
  ✓ WASH3000    →  largest car wash network in Vietnam
  ✓ Tasco Auto  →  6 service centers, roadside rescue, OBD diagnostics
  ✓ Carpla      →  used car marketplace
  ✓ VETC Wallet →  payments, balance, transaction history
  ✓ Insurance partner network

Every piece is there. The missing layer is intelligence.

Slide 4 — The Solution

VETC AI Agent Platform — a multi-agent AI system that unifies every car service into a single conversational interface.

One message. Six expert agents working together.

User: "Vừa chạy cao tốc xong, rửa xe gần nhất đi, và ví đủ tiền không?"

→ Trip Agent:   Hành trình HN→HP: 175,000₫ phí cầu đường. +150 điểm.
→ Wash Agent:   WASH3000 Mỹ Đình có slot 16:00, 85,000₫.
→ Wallet Agent: Ví còn 420,000₫ — đủ. Trừ tự động?
→ Response:     "Đặt rửa xe lúc 16:00 nhé? Tổng 85,000₫ từ ví VETC, bạn nhận thêm 200 điểm."

Slide 5 — Product Demo

6 Specialist Agents, 1 Interface

Agent What it does Example query
🛣️ Trip Toll costs, route history, spending patterns "Tháng này tôi tốn bao nhiêu phí đường?"
🚿 Wash WASH3000 slots, booking, subscriptions "Đặt rửa xe gần nhà tôi"
💰 Wallet Balance, top-up, transaction alerts "Nạp 500k vào ví VETC"
⭐ Loyalty Points, tiers, promotions "Tôi còn bao nhiêu điểm? Đổi được gì?"
🚨 Fine Traffic violation lookup, payment guidance "Biển 29A-12345 có bị phạt nguội không?"
🔧 Maintenance Warning lights, OBD codes, garage booking, rescue "Đèn cá vàng sáng, tôi phải làm gì?"

Execution Patterns

Pattern A — Single agent (default): One question → one expert answers in <2s.

Pattern B — Sequential chain: Fine Agent checks violations → Wallet Agent confirms if balance covers payment → merged answer.

Pattern C — Parallel fan-out: Post-trip event fires all agents simultaneously → single push notification with trip summary, wash suggestion, balance, and points earned.

Pattern E — Proactive nudge: Wallet drops below 100k₫ → agent pushes alert before user hits a toll booth.

Slide 6 — Market Opportunity

Vietnam Automotive Market

Signal Data
Registered vehicles 4.5M+ cars, growing 8% YoY
VETC e-tag installs 2M+ vehicles (mandatory for all highways)
WASH3000 addressable ~5M car washes/month market in Vietnam
Traffic fine market 10M+ fines issued/year via Nghị định 100/2019
Car maintenance market $2B+ annually

VETC as Distribution

The e-tag is already in 2M cars. No cold-start problem. No user acquisition cost for the core base.

No competitor can replicate this captive distribution.

Slide 7 — Business Model

Outcome-based pricing — charge per transaction facilitated, not per message or token.

Benchmarked against Sierra AI (~$1.50/resolution at $100M ARR):

Transaction Type Price per Event Volume Target (Year 1)
Wash booking via agent ~35,000₫ (~$1.40) commission 500 bookings/week
Service appointment booked ~50,000₫ (~$2.00) commission 200 bookings/week
Wallet top-up facilitated ~5,000₫ (~$0.20) per top-up 2,000/week
Traffic fine payment guided ~20,000₫ (~$0.80) per case 300/week
Insurance lead generated ~150,000₫ (~$6.00) per lead 50/week

Year 1 Revenue Target: ~5B₫/month (~$200K/month) at scale Unit economics: >70% gross margin (no physical costs, AI inference ~$0.02/query)

Future: Platform Licensing

Phase 3 opens the MCP server layer to 3rd parties (parking lots, petrol stations, EV charging). Platform fee: 15% of transactions facilitated through the agent network.

Slide 8 — Traction

Built in Sprint 0. Working today.

✅ 6 specialist agents — fully functional
✅ 5 mock MCP servers — Vietnamese data (Nghị định 100, OBD-II, WASH3000, Tasco Auto)
✅ 4 execution patterns — A, B, C, E
✅ Cross-agent unified memory — episodic store, 24h TTL, context-aware responses
✅ Streaming UI — mobile-first, SSE, markdown rendering
✅ 33/33 automated test cases passing — AI judge (Qwen) validates responses
✅ Vercel deployment ready

Test Suite Results (last run)

Group     Cases   Result
────────────────────────
Trip        3     3/3 ✓
Wallet      3     3/3 ✓
Wash        3     3/3 ✓
Loyalty     3     3/3 ✓
Fine        3     3/3 ✓
Maintenance 14    14/14 ✓
Pattern B   3     3/3 ✓
────────────────────────
TOTAL      33     33/33 (100%)

Slide 9 — Technology

Stack

Layer Choice Why
LLM Qwen2.5-14b-instruct Fastest response (1.2s), native Vietnamese, OpenAI-compatible API
Backend FastAPI + SSE Streaming responses, async tool execution
Agent framework Custom tool-use loop Full control, no framework lock-in
Tool protocol Mock MCP servers Production-ready swap: real APIs in same interface
Memory Episodic store (24h TTL) Cross-agent context without user repeating themselves
Frontend Vanilla JS Zero-dependency, mobile-first, instant load

Architecture Principles (from Sierra + Decagon playbook)

  1. Agents own domains, not conversations — orchestrator routes, agents execute
  2. MCP for tools, A2A for agents — loose coupling, independent deployment
  3. Outcome-based metrics — bookings and payments, not messages
  4. Data flywheel — every interaction enriches the user profile → better recommendations → more engagement

LLM Strategy

Don't lock to one model. Use Qwen for classification (cost + Vietnamese), Claude/GPT-4o for complex reasoning (quality), fine-tuned Qwen for domain-specific tasks (accuracy).

This is exactly Decagon's "model constellation" approach — the same pattern that got them to $1.5B valuation.

Slide 10 — Competitive Moat

What Exists Today

Company Valuation ARR What they built
Sierra AI $10B $100M Enterprise AI agent platform
Decagon $1.5B $35M Customer support agent OS
CrewAI — — Open-source multi-agent framework

All of them are platform-only. None of them own the services.

What VETC AI Has That They Can't Copy

① DATA + SERVICE OWNERSHIP (unique)
   Tasco owns the data layer (e-tag = trips, spending, mileage)
   AND the service supply (WASH3000, Tasco Auto, Carpla, Insurance)
   → Sierra/Decagon are middleware. We're the whole stack.

② CAPTIVE DISTRIBUTION (2M vehicles, zero CAC)
   VETC e-tag is legally required on every highway vehicle in Vietnam.
   The install base is already there. No cold start. No user acquisition cost.
   → No competitor can replicate this. It's government-mandated distribution.

③ COMPOUND CHAIN VALUE (defensible flywheel)
   Trip data → better Wash recommendations
   Wash frequency → better Maintenance predictions
   Maintenance history → better Insurance pricing
   Insurance savings → platform stickiness
   → Each domain makes every other domain smarter. No single competitor touches all.

④ VIETNAMESE MARKET SPECIFICS (local moat)
   Qwen fine-tuned on Vietnamese = language moat
   Nghị định 100/2019 fine database = regulatory moat
   Tasco's physical presence (service centers) = offline-online moat
   → A US AI company cannot replicate this in 12 months.

The AI is the glue. The ecosystem is the moat.

Slide 11 — Go-to-Market

Phase 1: Internal Beta (Month 1-2)

  • Deploy to Tasco employees + VETC power users (top 1% by trip frequency)
  • Focus: Wash booking and wallet top-up conversion
  • Measure: Agent resolution rate, booking conversion, NPS

Phase 2: Soft Launch (Month 3-4)

  • 10% of VETC user base (~200,000 users)
  • Proactive nudges enabled (Pattern E) — low balance, post-trip wash
  • A/B test: users with agent vs without
  • Target: 15% post-trip wash conversion, +25% auto-topup activation

Phase 3: Full Launch + Platform (Month 6-12)

  • All 2M+ VETC users
  • Open MCP layer to 3rd-party services (petrol stations, parking)
  • Insurance Agent + Marketplace Agent live
  • Platform licensing to other fleet operators

Slide 12 — Success Metrics

Metric Baseline 6-Month Target
Daily active users ~X% of VETC base +40%
Wash bookings via agent 0 500/week
Auto-topup activation ~Y% +25%
Avg agent response time N/A <2s
Agent resolution rate (no human) N/A >70% (Decagon benchmark)
Post-trip wash conversion 0% 15%
User NPS delta baseline +15 points
Revenue per user per month toll-only +30%

Slide 13 — Roadmap

NOW (Sprint 0) ✅
  6 agents, 5 MCPs, 4 patterns, memory, streaming UI, test suite

SPRINT 1 (next 4 weeks)
  Real VETC wallet API integration
  Real WASH3000 booking API integration
  Server-side Redis session store (replace client-carried history)
  Watchtower QA — auto-flag low-confidence wallet mutations

SPRINT 2 (Month 2)
  Insurance Agent + Marketplace Agent (Carpla trade-in, VETC Go rental)
  Pattern D — escalation chain (Maintenance → Insurance → Marketplace)
  Voice channel (in-car usage while driving)
  Metrics dashboard

SPRINT 3 (Month 3-4)
  LangGraph migration for stateful workflows + checkpointing
  A2A event bus (Redis Streams) for loose agent coupling
  Fine-tuned Qwen classifier on Tasco intent data
  OBD-II hardware integration for predictive maintenance

SPRINT 4 (Month 5-6)
  A/B launch to 10% of VETC users
  Open platform: 3rd-party MCP server registration
  Agent Studio: ops team tunes agents without code deploy

Built With

Share this project:

Updates