Evidence-backed multi-agent trust investigation system for companies, crypto projects, and wallets - powered by Vertex AI Gemini, Google ADK, MongoDB Atlas, and MongoDB MCP.
| Surface | Link | What judges can verify |
|---|---|---|
| Hosted product | https://vartovii-trust-agent-n7kszqvpoq-ew.a.run.app | Live Cloud Run product UI |
| Readiness proof | https://vartovii-trust-agent-n7kszqvpoq-ew.a.run.app/api/readiness | Gemini model, MongoDB status, Agent Engine path, MCP setup, test count |
| Judge trace | https://vartovii-trust-agent-n7kszqvpoq-ew.a.run.app/api/judge-trace | Agent route, evidence, decision path, audit events, MongoDB proof |
| Live CoinGecko proof | https://vartovii-trust-agent-n7kszqvpoq-ew.a.run.app/api/live-proof?slug=ethereum | Live market evidence plus MongoDB Atlas cache persistence |
| Live Etherscan proof | https://vartovii-trust-agent-n7kszqvpoq-ew.a.run.app/api/wallet-live-proof | Live ETH wallet balance proof plus MongoDB Atlas cache persistence |
| Open-source repo | https://github.com/Vetassikc/vartovii-trust-agent | MIT license, source code, docs, tests, deploy scripts |
Trust work usually fails in two ways: data is scattered, and AI answers are hard to verify. A user may need to check market APIs, blockchain explorers, company records, search results, and internal notes before making one trust decision.
Vartovii is built around visible investigation, not a black-box answer. A judge can verify that the system:
- routes work through specialist Google ADK agents;
- uses Gemini 3.5 Flash as the active production model;
- pulls live CoinGecko and Etherscan evidence;
- stores investigations, audit events, and live proof in MongoDB Atlas;
- exposes proof through readiness, judge trace, live evidence, and wallet proof endpoints;
- uses MongoDB MCP as a flexible database inspection path beyond fixed tools.
The result is an evidence-first trust console: route, score, sources, persistence, and decision in one inspectable workflow.
This repository is packaged as a standalone contest build created for the Google Cloud Rapid Agent Hackathon. The domain comes from trust intelligence research patterns, while the submitted agent graph, proof endpoints, UI, tests, deployment scripts, MongoDB Atlas integration, and MongoDB MCP path are focused on this judge-verifiable hackathon implementation.
graph TB
User([🧑 User Query]) --> Orchestrator
subgraph "Vartovii ADK Multi-Agent System"
Orchestrator["🎯 Orchestrator Agent<br/>Gemini 3.5 Flash<br/><i>Pure delegation — never answers directly</i>"]
Orchestrator -->|"Company questions"| Corporate["🏢 Corporate Agent<br/>6 FunctionTools"]
Orchestrator -->|"Crypto questions"| Crypto["🪙 Crypto Agent<br/>6 FunctionTools"]
Orchestrator -->|"Web research"| OSINT["🔍 OSINT Agent<br/>GoogleSearchTool"]
Orchestrator -->|"Save & recall"| Memory["🧠 Memory Agent<br/>4 FunctionTools"]
Orchestrator -->|"Ad-hoc DB questions"| MCPAgent["🍃 MongoDB MCP Agent<br/>(optional)"]
MCPAgent --> MCP["🔌 MongoDB MCP Server"]
end
Corporate --> DB[(MongoDB Atlas)]
Crypto --> DB
Memory --> DB
MCP --> DB
Corporate --> CT1[search_company]
Corporate --> CT2[get_trust_score]
Corporate --> CT3[compare_companies]
Corporate --> CT4[list_companies]
Corporate --> CT5[get_company_reviews]
Corporate --> CT6[get_vacancy_intelligence]
Crypto --> CRT1[search_crypto_projects]
Crypto --> CRT2[get_crypto_trust_score]
Crypto --> CRT3[check_wallet]
Crypto --> CRT4[get_transaction_history]
Crypto --> CRT5[get_token_holders]
Crypto --> CRT6[get_contract_info]
Memory --> M1[save_investigation]
Memory --> M2[get_investigation_history]
Memory --> M3[log_audit_event]
Memory --> M4[get_audit_trail]
OSINT --> GS[Google Search Grounding]
style Orchestrator fill:#1a73e8,stroke:#1557b0,color:#fff
style Corporate fill:#34a853,stroke:#2d8e47,color:#fff
style Crypto fill:#ea8600,stroke:#c77200,color:#fff
style OSINT fill:#9334e6,stroke:#7b2bc1,color:#fff
style Memory fill:#e8453c,stroke:#c23a32,color:#fff
style MCP fill:#47A248,stroke:#3d8b3d,color:#fff
style DB fill:#47A248,stroke:#3d8b3d,color:#fff
5 core agents, 28 custom tools, and an optional MongoDB MCP specialist — orchestrated by Google ADK for autonomous trust intelligence.
| Feature | Description |
|---|---|
| 🤖 Multi-Agent Orchestration | 5 specialized LlmAgents coordinated via Google ADK — each with its own tools, context, and domain expertise |
| 🍃 MongoDB Atlas + MCP | Structured PyMongo tools handle production workflows; the optional mongodb-mcp-server specialist handles ad-hoc collection inspection, aggregation, and explain-plan work |
| 📡 Live Evidence Proof | /api/live-proof fetches CoinGecko market evidence, while /api/wallet-live-proof fetches Etherscan ETH balance evidence and persists both proof paths in MongoDB Atlas cache |
| 🔄 Model Fallback | Production uses Vertex AI Gemini 3.5 Flash with a 3.1 Flash-Lite cost profile and explicit 3.1 Pro preview opt-in |
| 🧠 Investigation Memory | Cross-session persistence: agents save & recall past investigations via MongoDB |
| 📋 Audit Events | Investigation and proof actions are recorded with agent, action, model, latency, and timestamp metadata |
| 🔍 OSINT Grounding | Real-time web research via Google Search Grounding for entities not in database |
| 📊 28 Specialized Tools | Corporate analytics, crypto forensics, wallet checks, on-chain analysis, similarity search, network risk, investigation management |
| 🧪 63 Automated Tests | Architecture validation, MCP construction, live proof contracts, dashboard fallback/readiness behavior, model routing, service layer coverage |
- Python 3.11+
- Google API Key
- MongoDB Atlas cluster (free tier works)
- Node.js 18+ (for MongoDB MCP Server)
# Clone the repository
git clone https://github.com/Vetassikc/vartovii-trust-agent.git
cd vartovii-trust-agent
# Create virtual environment
python -m venv .venv && source .venv/bin/activate
# Install dependencies
pip install -e '.[dev]'
# Configure environment
cp .env.example .env
# Fill in GOOGLE_API_KEY, MONGODB_CONNECTION_STRING, and optional ETHERSCAN_API_KEYpython scripts/seed_mongodb.pyThis populates your MongoDB Atlas cluster with:
- 🏢 Corporate entities with trust scores, reviews, vacancy data
- 🪙 Crypto projects with tokenomics, on-chain metrics, holder distributions
- 🔗 Wallet records and transaction histories
To restore the non-destructive judge proof path without dropping core collections, run:
python scripts/seed_judge_evidence.pyThis upserts the Wirecard judge investigation, replayable audit events, and normalizes legacy audit model labels to the active Gemini policy.
# ADK Web Interface (recommended for demo)
adk web agent/
# Or run interactive demo scenarios
python -m demo.run_demo# Web console + FastAPI + MongoDB MCP child process
./scripts/deploy.sh
# ADK agent graph on Google Cloud Agent Engine
GOOGLE_CLOUD_PROJECT=your-project ./scripts/deploy_agent_engine.shCloud Run is the primary hosted product demo because the container includes the
dashboard API, static web console, and Node.js runtime for mongodb-mcp-server.
Agent Engine deployment proves the ADK agent graph is ready for Google Cloud's
hosted agent runtime; set MONGODB_MCP_ENABLED=true only in runtimes where the
MongoDB MCP child process is available.
The Agent Engine deploy helper uses a temporary sanitized env file by default: it keeps MCP disabled and uses mock data fallback so the hosted graph can be deployed without copying local secrets. The Cloud Run deployment remains the live MongoDB + MCP product surface.
Vartovii connects to MongoDB Atlas through two complementary pathways:
Each agent uses purpose-built FunctionTools that query MongoDB collections through a singleton connection manager (agent/tools/db.py):
# Example: Corporate Agent's search_company tool
from agent.tools.db import get_collection
collection = get_collection("companies")
result = collection.find({"name": {"$regex": query, "$options": "i"}})The official mongodb-mcp-server runs as a subprocess, exposing MongoDB operations via the Model Context Protocol:
# Initialized in agent/adk_agent.py as an optional MongoDB MCP specialist
toolset = McpToolset(
connection_params=StdioConnectionParams(
server_params=StdioServerParameters(
command="npx",
args=["-y", "mongodb-mcp-server"],
env={"MONGODB_CONNECTION_STRING": conn_string},
),
timeout=10.0,
),
)This gives the ADK graph direct, flexible database access through a dedicated MCP specialist — it can run ad-hoc finds, aggregations, and explain plans without needing a pre-built tool for every query pattern.
| Collection | Purpose | Key Fields |
|---|---|---|
companies |
Corporate entity data | name, trust_score, country, industry, reviews |
crypto_projects |
Crypto project profiles | name, symbol, trust_score, tvl, security_score |
wallets |
Blockchain wallet records | address, chain, balance, transactions |
investigations |
Saved investigation results | entity_name, entity_type, trust_score, risk_level, timestamp |
audit_log |
Agent action audit trail | agent, action, model_used, latency_ms, timestamp |
| Layer | Technology |
|---|---|
| AI Framework | Google Agent Development Kit (ADK) 2.0.0 |
| Models | Vertex AI Gemini 3.5 Flash GA; Gemini 3.1 Flash-Lite cost profile; Gemini 3.1 Pro preview opt-in |
| Database | MongoDB Atlas — cloud-hosted document database |
| MCP Server | mongodb-mcp-server — official MongoDB MCP integration |
| Driver | PyMongo 4.7+ with connection pooling & retry |
| Language | Python 3.11+ |
| Search | Google Search Grounding (OSINT agent) |
| Testing | pytest 8.0+, pytest-asyncio |
| Deployment | Google Cloud Run web demo + ADK Agent Engine deployment path + MongoDB Atlas |
vartovii-trust-agent/
├── agent/ # Core ADK agent definitions
│ ├── __init__.py # Package init (exports root_agent)
│ ├── agent.py # ADK entry point (symlink to adk_agent.py)
│ ├── adk_agent.py # Root orchestrator + 4 sub-agents + MCP
│ ├── requirements.txt # Agent Engine packaging dependencies
│ ├── config.py # Model routing, fallback chains, MongoDB config
│ ├── prompts/
│ │ └── adk.py # Agent instruction prompts
│ └── tools/
│ ├── corporate_tools.py # 6 corporate intelligence tools
│ ├── crypto_tools.py # 6 crypto forensics tools
│ ├── investigation_tools.py # 4 investigation & audit tools
│ ├── db.py # MongoDB connection manager (singleton)
│ ├── live_data.py # CoinGecko and Etherscan live evidence helpers
│ └── mock_data.py # Fallback demo data providers
├── services/ # Service layer
│ ├── model_runtime.py # Model execution with fallback chain
│ ├── routing_adapter.py # Chat routing adapter
│ └── telemetry.py # Metrics and monitoring
├── scripts/
│ ├── deploy.sh # Cloud Run web demo deploy
│ ├── deploy_agent_engine.sh # ADK Agent Engine deploy
│ └── seed_mongodb.py # Seed MongoDB Atlas with demo data
├── tests/
│ ├── test_agent.py # 38 agent architecture & tool tests
│ ├── test_dashboard_api.py # 12 dashboard fallback/readiness/live-proof contract tests
│ └── test_services.py # 13 service layer tests
├── demo/
│ └── run_demo.py # Interactive demo runner (5 scenarios)
├── web/ # Dashboard frontend
│ ├── index.html # Main UI
│ ├── style.css # Styles
│ └── app.js # Frontend logic
├── evidence/ # Hackathon submission evidence
│ ├── optimization_metrics.md # Before/after metrics
│ ├── production_rollout_report.md
│ └── screenshots/
├── AGENTS.md # Repository rules for coding agents
├── AGENT_ROLE_MAPPING.md # Agent ownership, tools, and handoff rules
├── MODEL_POLICY.md # Model routing, fallback, and preview policy
├── PROJECT_CONTEXT.md # Product context and judging narrative
├── SOURCE_UPDATE_POLICY.md # Evidence freshness and source governance
├── pyproject.toml # Project config & dependencies
├── .env.example # Environment variable template
├── ARCHITECTURE.md # Detailed technical architecture
├── LICENSE # MIT License
└── README.md # ← You are here
# Run all tests
pytest tests/ -v
# Run with coverage
pytest tests/ -v --tb=short| Test Suite | Tests | Coverage |
|---|---|---|
test_agent.py |
38 | Agent topology, tool registration, MCP integration, fallback chains, audit model policy |
test_dashboard_api.py |
12 | Dashboard mock fallback, readiness endpoint, CoinGecko and Etherscan live proof contracts, judge trace, health model metadata, leaderboard and entity detail contracts |
test_services.py |
13 | Model runtime, routing adapter, telemetry, config validation |
| Total | 63 | Architecture, tools, dashboard API, MongoDB fallback/readiness/live proof, services |
Key test categories:
- ✅ Agent architecture — verifies 5-agent topology, correct tool assignment
- ✅ Tool contracts — validates custom tools and dashboard API contracts return expected schemas
- ✅ MongoDB integration — connection manager, collection access, graceful fallback
- ✅ Model routing — profile switching and resilient fallback chains
- ✅ MCP toolset — initialization, error handling, connection params
> "Analyze SAP as an employer"
→ Orchestrator → Corporate Agent
→ search_company("SAP") → get_trust_score("SAP")
→ Trust Score: 74/100 | Risk: MEDIUM | 6-pillar breakdown
→ Memory Agent saves investigation to MongoDB
> "Give me the full trust assessment for Uniswap"
→ Orchestrator → Crypto Agent
→ search_crypto_projects("Uniswap") → get_crypto_trust_score("uniswap")
→ Trust Score: 78/100 | Security: HIGH | TVL, dev activity, audit status
→ Memory Agent saves investigation to MongoDB
GET /api/live-proof?slug=ethereum
→ Crypto proof path fetches CoinGecko market evidence
→ Trust delta is computed from 24h movement
→ MongoDB Atlas caches the live evidence for audit and MCP inspection
→ Response exposes source URL, freshness, persistence status, and agent trace
GET /api/wallet-live-proof
→ Wallet proof path fetches native ETH balance from Etherscan API V2
→ Balance evidence is normalized for judge-readable proof cards
→ MongoDB Atlas caches the wallet evidence for audit and MCP inspection
→ Response exposes source URL without apikey, freshness, persistence status, and agent trace
> "Check wallet 0xd8dA6BF26964aF9D7eEd9e03E53415D37aA96045"
→ Orchestrator → Crypto Agent
→ check_wallet("0xd8dA...") → get_transaction_history("0xd8dA...")
→ Live balance from Etherscan when configured | Recent transactions | Risk flags
> "Show me all past crypto investigations"
→ Orchestrator → Memory Agent
→ get_investigation_history(entity_type="crypto")
→ List of past investigations with scores and timestamps
Primary task model ──on error──▶ Task fallback when different ──on error──▶ Ultimate fallback
| Profile | Agent Model | Chat Model | Report Model |
|---|---|---|---|
| stable | gemini-3.5-flash |
gemini-3.5-flash |
gemini-3.5-flash |
| cost | gemini-3.1-flash-lite |
gemini-3.1-flash-lite |
gemini-3.1-flash-lite |
| preview | gemini-3.5-flash |
gemini-3.5-flash |
gemini-3.1-pro-preview |
All models are environment-overridable. The resolved chain deduplicates repeated
model names, so stable chat and agent paths currently resolve to
gemini-3.5-flash -> gemini-2.0-flash. The fallback policy reduces
user-visible model failures without hiding routing metadata.
MIT — Vitalii Radionov, 2026
🏆 Built for Google Cloud Rapid Agent Hackathon — MongoDB Track
Autonomous multi-agent trust intelligence powered by Vertex AI Gemini, Google ADK, and MongoDB Atlas via MCP