Skip to content

DEMO_SCRIPT.md

Latest commit

 

History

History
166 lines (117 loc) · 7.78 KB

DEMO_SCRIPT.md

File metadata and controls

166 lines (117 loc) · 7.78 KB

Aura — Hackathon Demo Script

Event: Google AI Agent Labs Oslo 2026
Team: Team 6
Total time: ~5 minutes pitch + 3 minutes live demo + 2 minutes Q&A

Elevator Pitch (30 seconds)

"We built Aura — Autonomous Reliable Agentic Commerce. It's a multi-agent AI system that automates the entire B2B procurement lifecycle: vendor discovery, KYC/AML compliance, and payment settlement — all from a single natural language request. What makes Aura unique is compliance-first design: we check every vendor against the compliance database before generating a payment mandate, not after. No human in the loop, no compliance gaps."

Part 1 — Problem (60 seconds)

Talking points:

  • B2B procurement today is manual, slow, and fragmented — sourcing, compliance, and payments are three separate silos
  • Most AI "shopping bots" skip compliance entirely — they find a vendor and pay. That's a financial crime risk
  • Banks and enterprises are under increasing AML/KYC regulatory pressure
  • The gap we're closing: AI agents that can transact autonomously, with built-in compliance proof on every single transaction

One-liner to drop:

"ShadowHardware looks cheap. But it's on the AML blacklist. Traditional bots would pay them. Aura blocks them — automatically, every time."

Part 2 — Architecture Walkthrough (60 seconds)

Point to the architecture diagram (docs/ARCHITECTURE.md) or draw the flow on the whiteboard:

User → Architect → Scout → Sentinel → Closer → Settlement
                              ↓ BLOCKED
                         No payment ever made

Key points per agent:

Agent One sentence
Architect Root orchestrator — parses your intent, manages the squad, summarises the outcome
Scout Queries the UCP vendor network — finds ALL vendors, flags risky ones, never filters
Sentinel Compliance gate — KYC/AML check on EVERY vendor via the Core Banking system. One REJECTED = full block
Closer Generates a signed W3C Verifiable Credential (Intent Mandate) and submits to AP2 gateway — only if Sentinel approved

Tech stack drop:

"Built on Google ADK with Gemini 2.5 Flash via Vertex AI. Deployed as Kubernetes agents using Kagent v1alpha2 CRDs — so it's cloud-agnostic. Swap Gemini for Bedrock or Azure OpenAI by changing two lines of YAML."

Part 3 — Live Demo (3 minutes)

Setup (before presenting — run these in advance)

# Terminal 1 — start the dashboard
cd /path/to/aura
source .venv/bin/activate
streamlit run ui/dashboard.py
# Opens at http://localhost:8501

# Terminal 2 — optionally start the API server too
uvicorn main:app --reload --port 8080

Demo Step 1 — Happy Path (90 seconds)

Open the dashboard at http://localhost:8501.

  1. Type in the input box: 
    Buy 3 Laptop Pro 15 units from the best available vendor
  2. Click Launch Procurement
  3. Narrate as each agent card activates:
    • 🏛️ Architect: "Intent parsed — delegating to pipeline"
    • 🔭 Scout: "Discovers 4 vendors — including ShadowHardware at $899, suspicious but included, that's the Scout's job"
    • 🛡️ Sentinel: "Runs KYC/AML on all 4 — watch ShadowHardware get flagged. The other three pass with Compliance Hashes"
    • 💳 Closer: "Selects NordHardware at $1,280 — cheapest approved vendor. Generates an Intent Mandate, signs it, settles via AP2"
  4. Point to the settlement card:

    "Settlement confirmed — $3,840, Settlement ID AP2-XXXX. And you can see ShadowHardware was excluded automatically. No one had to make a judgment call."

Demo Step 2 — Compliance Block (60 seconds)

  1. Clear the input and type: 
    Buy laptops from ShadowHardware
  2. Click Launch Procurement (or use the Blocked Path quick button)
  3. Narrate:
    • Scout finds ShadowHardware
    • Sentinel hits the BMS compliance check — REJECTED: AML_BLACKLIST
    • Closer card turns red: PAYMENT_ABORTED — zero payment tools called
  4. Key line:

    "The Closer never even saw a payment request. The Compliance Hash was never generated. There is no way for money to move to a blacklisted vendor — it's architectural."

Optional Demo — API (30 seconds, if time permits)

curl -X POST http://localhost:8080/run \
  -H "Content-Type: application/json" \
  -d '{"message": "Buy 3 Laptop Pro 15 units from best vendor"}'

"It's also a REST API — so any enterprise system can plug in. Multi-turn sessions, streaming responses, Kubernetes-native."

Part 4 — Key Differentiators (30 seconds)

Three things that make Aura different:

  1. Compliance-first, not compliance-after — the Compliance Hash is embedded in the payment mandate. The gateway verifies it independently. No compliance, no payment — architecturally enforced.

  2. Verifiable payment mandates — W3C Verifiable Credential structure with ECDSA-P256 proof. Every transaction is cryptographically auditable.

  3. Cloud-agnostic, Kubernetes-native — Kagent CRDs means each agent scales independently and runs on GCP, AWS, Azure, or on-premise.

Anticipated Judge Questions

Q: This is a mock — how close is it to production?

"The agent logic, pipeline, session state handoff, and data structures are all production-grade. The three integrations that are mocked — UCP vendor discovery, BMS compliance API, and AP2 settlement — are all designed as drop-in replacements. The tools/ layer is fully decoupled from the agent reasoning layer. Replacing mock with real is a single function swap per tool."

Q: What's to stop the LLM from ignoring the Sentinel's output?

"Two layers of safety. First, the Closer's system prompt has an explicit hard override: if COMPLIANCE_BLOCKED appears anywhere in session state, output PAYMENT_ABORTED and call no tools. Second, settle_cart_mandate() validates the compliance_hash field before processing — no hash means an exception is raised. The LLM can't accidentally bypass it."

Q: How does session state pass between agents?

"Google ADK's SequentialAgent passes session state automatically. Each agent writes to its output_key — scout_results, sentinel_results, closer_results. The next agent reads the full accumulated context including those keys in its prompt window."

Q: Can it handle multiple procurement requests concurrently?

"Yes — each /run call gets its own session_id and ADK session. The InMemorySessionService handles concurrent sessions. For production scale, we'd swap in a persistent session store like Firestore."

Q: Why Kagent?

"Kagent gives us cloud-agnostic Kubernetes-native agent deployment. Each agent is a CRD, scales independently, and can run on any CNCF-conformant cluster. Swapping the model backend is two lines of YAML — we support Vertex AI, AWS Bedrock, Azure OpenAI, and Ollama out of the box."

Q: What's the $5,000 transaction cap?

"That's a deliberate financial control built into generate_intent_mandate(). The Intent Mandate's constraints field enforces max_amount: 5000.00. This mirrors how real AP2 mandates work — the mandate embeds the constraint, and the gateway validates it before processing."

Closing Line

"Aura is what B2B commerce looks like when agents are first-class participants — not just bots that scrape prices, but autonomous actors that discover, vet, and pay with cryptographic proof at every step. Thank you."

Quick Reference Links

What Where
Dashboard http://localhost:8501
API Swagger UI http://localhost:8080/docs
Architecture diagram docs/ARCHITECTURE.md
Agent flow diagrams docs/AGENT_FLOW.md
Business Guide docs/BUSINESS_GUIDE.md
Technical Guide docs/TECHNICAL_GUIDE.md
⚡