HowDidi
HowDidi
⚡ Submission Hub · Google Cloud Rapid Agent Hackathon

HowDidi — Agentic Operations
for Modern Ecommerce

A real-world operator that ingests ecommerce data via **Fivetran MCP**, diagnoses revenue leaks using **Vertex AI Reasoning Engines**, and drafts action-ready fixes for one-tap owner approvals.

⚡ Quick Judge Actions

Select the pre-seeded Streetwear Store demo and jump straight to the core experience in under 60 seconds:

🏗️ Unified Stack & Data Flow

HowDidi is a secure, production-grade hybrid service separating analytical operations from secure GCP execution.

1. Client & Server

React & TanStack Start

A premium, responsive UI served on managed Google Cloud Run. Leverages SSR data loaders and real-time state synchronizations. Fully styled under the Dark-Neon Night Market design system.

2. AI Agent Core

Vertex AI Reasoning Engine

Registers our Python ADK agent (`infra/adk-agent/agent.py`) as a GCP Vertex Reasoning Engine. Securely connects with standard service accounts and runs dynamically compiled telemetry payloads.

3. Data layer

Fivetran MCP Server

Provides a secure read-only API connecting Shopify, Stripe, Meta Ads, and Help Scout. When live credentials are not set, a local SQLite seed database populates full storefront profiles dynamically.

📋 Judging Criteria Mapping

1. Technological Implementation

  • Vertex AI ADK / Reasoning Engine deeply integrated inside python agent stack (`infra/adk-agent/agent.py`)
  • Dynamic programmatic registration & live gcloud invocation
  • Fivetran Model Context Protocol (MCP) data ingestion layer with UVX runner fallback
  • Robust OpenTelemetry instrumentation linking all agent spans directly to Google Cloud Trace

2. Design & UX

  • Premium 'Night Market' dark-neon command operator aesthetic
  • Highly responsive dual code paths ensuring seamless navigation across both mobile and desktop screen heights
  • Cinematic Mission HUD console tracking agent reasoning feeds, execution times, and tools invocations in real-time

3. Potential Impact

  • Explicitly quantifies revenue leak sums ($) across abandoned carts, stockouts, ad waste, and urgent tickets
  • Pre-composes actionable drafts (such as Shopify support replies or paused ad sets) ready to run in one-tap
  • Saves store owners hundreds of hours of daily manual cross-tab spreadsheet lookups

4. Quality of Idea

  • Autonomous operations center model — NOT another passive read-only SaaS dashboard
  • Human-in-the-loop validation paradigm: Didi prepares the work, the owner approves and fires the fix
🔒 Security & Privacy Commitments

Built for Approval-Safe Operations

Autonomous store agents must never execute writes directly without verification. HowDidi enforces absolute safety:

  • ✓ 100% Read-Only connectors (`FIVETRAN_ALLOW_WRITES=false`)
  • ✓ Human-in-the-loop manual action gates
  • ✓ Decoupled tokens managed entirely via GCP Secret Manager
  • ✓ Production sandbox environment isolation

🧪 Diagnostic Automated Evals

Every single agent configuration (Mock, Gemini 2.0, Python ADK) is continuously evaluated against 8 rigid operational metrics:

Adapter Type
Checks Passed
Status
Deterministic Mock Adapter
8 / 8 Checkpoints
🟢 100% Correct
Direct Gemini Adapter
8 / 8 Checkpoints
🟢 100% Correct
Python ADK Reasoning Adapter
8 / 8 Checkpoints
🟢 100% Correct
Checkpoints verified by the suite:
1.Health Score Range Verification: Checks that overall pulse scores are mapped exactly within [40, 75]
2.Checkout & Cart Focus Analysis: Ensures the highest-value abandoned carts are ranked as top priority leaks
3.Severity Assessment: Validates that high-severity stockouts or ad wastes are flagged with high/critical levels
4.Revenue Impact Pricing: Verifies that estimated leak impacts are mathematically scaled above $500
5.Action Generation Cardinality: Checks that at least 4 individual actionable recommendations are drafted
6.Safety Approval Gates: Ensures that every action is strictly queued under requireApproval status
7.Comprehensive Synthesis Check: Validates that synthesized paragraphs contain rich context and full details
8.Evidence Attribution & Sources: Ensures that all telemetry leaks link back to correct source connectors
npm run eval

🔗 Essential Hackathon Resources

GitHub Code Repo •About the Agent •REST Health API •Agent Runtime API