Autonomous KYC/AML Intelligence Terminal — research any company, person, or fund and receive a structured financial crime risk report in under 4 minutes, powered by Google ADK · Gemini · MongoDB Atlas.
Live demo: hawk-ai-flax.vercel.app
Backend health: hawkai-backend-46616099912.us-central1.run.app/api/v1/health
Financial crime costs the global economy $2 trillion+ per year. Yet the tools to fight it are locked behind Bloomberg Terminal licenses at $24,000/year — inaccessible to 99% of compliance teams at community banks, fintech startups, and financial regulators.
A compliance officer manually screening a single counterparty spends 2–4 days and $100+ per entity. HawkAI does it autonomously in under 4 minutes for fractions of a cent.
Type a name. HawkAI does the rest.
A multi-agent pipeline autonomously searches the web for compliance intelligence, extracts a structured entity profile, runs 768-dimensional vector similarity search to surface correlated entities, classifies risk signals across 8 categories, and synthesizes a scored risk report — all streamed live to a Bloomberg Terminal-style UI.
- Risk score (0–100) with CRITICAL / HIGH / MEDIUM / LOW classification
- Confidence score (0–100%) reflecting source quality and coverage
- 5–7 specific findings, each citing verifiable facts (dollar amounts, dates, regulatory bodies)
- Compliance recommendations in PRIMARY ACTION / MONITORING / PERIODIC REVIEW structure
- Risk breakdown by category — GOVERNANCE · FRAUD · SANCTIONS · FINANCIAL · REGULATORY · LITIGATION · REPUTATION
- Entity Relationship Network — force-directed graph showing risk correlations across all investigated entities, connected by shared signal categories
- Batch screening — comma-separate multiple names to queue simultaneous investigations
- Watchlist alerts — star any entity; get a live badge count when watched entities hit HIGH/CRITICAL risk
- Live event feed showing every agent action, tool call, and pipeline step in real time
- PDF export — full multi-page report with one click
- Light / dark mode — persisted to localStorage
- Analyst notes — add and store observations against any entity profile
⚠️ KYC/AML compliance screening only. Scores reflect regulatory enforcement history and adverse findings, not investment grade or creditworthiness.
┌──────────────────────────────────────────┐
│ HawkAI Bloomberg Terminal UI │
│ Next.js 14 · TypeScript · App Router │
│ JetBrains Mono · amber-on-ink palette │
└──────────────┬───────────────────────────┘
│ REST + SSE streaming
┌──────────────▼───────────────────────────┐
│ FastAPI Backend │
│ /api/v1/* · SSE · Google Cloud Run │
│ min-instances=1 (no cold start) │
└──────────────┬───────────────────────────┘
│
┌──────────────▼───────────────────────────┐
│ ScoutOrchestrator │
│ Google ADK SequentialAgent │
└──────────┬────────────────────────────────┘
│
┌─────────────────────▼──┐ ┌───────────────────────────────────┐
│ ResearchAgent │ │ IntelligenceAgent │
│ gemini-2.5-flash │ │ gemini-2.5-flash │
│ google_search (live) │ │ + gemini-2.5-pro for synthesis │
│ 2 targeted searches │ │ 7 custom async Python tools │
└────────────────────────┘ └──────────────┬────────────────────┘
│
┌──────────────────┤
│ lookup_entity_via_mcp → MongoDB MCP Server (ADK MCPToolset)
│ check_ofac_sanctions → MongoDB MCP Server → sanctions_lists (17,557 SDN entries)
│ extract_and_store_entity → Motor async upsert
│ run_vector_similarity_search → $vectorSearch aggregation
│ find_correlated_entities → Motor async find
│ classify_and_store_signals → gemini-2.5-pro + Motor
│ synthesize_risk_report → gemini-2.5-pro + Motor
└──────────────────┬──────────────────
│
┌────────────────────────────────▼──────────────────────┐
│ MongoDB Atlas M0 │
│ investigations · entities · risk_signals │
│ sanctions_lists (OFAC SDN — 17,557 entries) │
│ Vector Search Index (768-dim cosine) │
│ gemini-embedding-001 embeddings │
└────────────────────────────────────────────────────────┘
| Decision | Why |
|---|---|
| SequentialAgent | ADK constraint: google_search cannot share a session with database tools. Two agents, one pipeline. |
| MongoDB MCP Server + Motor dual-driver | Two MCP tools run on every investigation: lookup_entity_via_mcp checks for existing profiles; check_ofac_sanctions screens against the pre-loaded OFAC SDN list (17,557 US Treasury sanctioned entities). Both use ADK MCPToolset + pre-installed mongodb-mcp-server. Motor handles all writes for atomic guarantees in Cloud Run's serverless environment. |
| Tiered Gemini models | gemini-2.5-flash for agent orchestration (fast, live search, tool calling); gemini-2.5-pro for signal classification + report synthesis (highest output quality). |
| Person disambiguation | Optional context field (company, role, country, year) sent with person investigations to identify the right individual among common names. |
| Evidence-weighted scoring | 60% max-severity floor + 40% weighted signal average + signal-breadth bonus. One CRITICAL signal guarantees a CRITICAL (75+) score. |
| Force-directed entity graph | Pure JavaScript simulation (Coulomb repulsion + Hooke springs + center gravity), 250 synchronous iterations — no D3 dependency, stable layout before first render. |
| SSE streaming | Every agent event (tool call, step, text, snapshot) is pushed to the frontend as it happens — the UI never waits on a response body. |
| Layer | Technology |
|---|---|
| Agent Framework | Google ADK 2.0 (SequentialAgent) |
| Agent Model | Gemini 2.5 Flash — live Google Search + tool calling |
| Synthesis Model | Gemini 2.5 Pro — signal classification + report generation |
| Embeddings | gemini-embedding-001 (768-dim) |
| Web Research | Google Search grounding (ADK built-in, live OSINT) |
| Database | MongoDB Atlas M0 (free tier) |
| DB Protocol | MongoDB MCP Server (mongodb-mcp-server) via ADK MCPToolset — entity lookup + OFAC SDN screening |
| DB Driver | Motor async driver — writes, vector search, signal storage |
| Vector Search | MongoDB Atlas $vectorSearch (cosine, 768-dim) |
| Backend | FastAPI + asyncio + SSE |
| Frontend | Next.js 14 + TypeScript + App Router |
| UI | Bloomberg Terminal-style — JetBrains Mono, amber phosphor palette |
| Deployment | Google Cloud Run (us-central1) + Vercel |
Every report includes an executive summary, key findings numbered with category tags (FRAUD · SANCTIONS · GOVERNANCE…), recommendations split by action type, and a risk breakdown sidebar with full category names and signal counts. Reports are exportable to PDF with one click.
Force-directed SVG graph connecting every investigated entity. Nodes are sized by risk score, colored by risk level (CRITICAL → red, HIGH → orange, MEDIUM → amber, LOW → green). Edges represent shared risk signal categories. Hover any node for a score/level tooltip; click to open the full report. Edge labels show the shared signal types on hover.
Comma-separate names in the query field: Binance, Tether, Circle. A ⇶ BATCH · 3 ENTITIES indicator appears. All investigations are created simultaneously and the pipeline streams the first one; the rest populate the dossier table as they complete in the background.
Star (★) any row in the dossier table. If a watched entity completes with HIGH or CRITICAL risk, a red badge appears on the WATCHLISTS nav item and the Watchlists view shows an ACTIVE ALERTS section at the top with one-click cards to open each flagged report.
Every pipeline step, tool call, and agent output is streamed in real time to the terminal feed and the live queue panel. Nothing is polled — pure SSE.
| Dossier Table | Investigation Report |
|---|---|
 |
 |
| Entity Risk Network | Risk Signals |
|---|---|
 |
 |
- Google AI Studio API Key (paid tier) → aistudio.google.com
- MongoDB Atlas M0 cluster (free) → mongodb.com/atlas
- Docker + Docker Compose, or Python 3.12+ / Node 22+
git clone https://github.com/Nstar9/HawkAI.git && cd HawkAI
cp backend/.env.example backend/.envEdit backend/.env:
GOOGLE_API_KEY=your-google-ai-studio-key
MONGODB_URI=mongodb+srv://user:pass@cluster.mongodb.net/hawkai
GEMINI_MODEL=gemini-2.5-flash
SYNTHESIS_MODEL=gemini-2.5-proIn Atlas UI → Atlas Search → Create Index → JSON Editor, collection entities:
{
"fields": [
{ "type": "vector", "path": "embedding", "numDimensions": 768, "similarity": "cosine" },
{ "type": "filter", "path": "entity_type" },
{ "type": "filter", "path": "risk_level" }
]
}docker-compose up --build
# → http://localhost:3000Verified live scores from the deployed system:
| Entity | Type | Score | Level | Notable Signals |
|---|---|---|---|---|
| FTX | COMPANY | ~84 | CRITICAL | Founder convicted on 7 counts, 25-year sentence, CFTC $12.7B consent order, Chapter 11 |
| Sam Bankman-Fried | PERSON | ~81 | CRITICAL | Wire fraud, securities fraud, money laundering — SDNY Docket 22-CR-00673 |
| Tornado Cash | COMPANY | ~82 | CRITICAL | OFAC-sanctioned, $7B+ laundered including for DPRK hackers |
| Vanguard | COMPANY | ~77 | HIGH | $146M SEC settlement (Jan 2025), regulatory enforcement history |
| Meta Inc | COMPANY | ~55 | HIGH | FTC consent orders, Cambridge Analytica, $5B privacy fine |
| Tesla Inc | COMPANY | ~48 | MEDIUM | Musk SEC securities violations, civil settlement, shareholder litigation |
| Fidelity Investments | COMPANY | ~28 | MEDIUM | Routine regulatory filings, no material enforcement history |
| BlackRock Inc | COMPANY | ~8 | LOW | Clean — world's largest asset manager, standard regulatory disclosures |
Score interpretation: 0–21 LOW (standard onboarding) · 22–51 MEDIUM (enhanced monitoring) · 52–79 HIGH (escalate for review) · 80–100 CRITICAL (reject / file SAR)
Scores are KYC/AML compliance signals — not investment advice and not legal verdicts.
| Method | Endpoint | Description |
|---|---|---|
GET |
/api/v1/health |
Service health — returns model config |
POST |
/api/v1/investigations |
Start a new investigation |
GET |
/api/v1/investigations |
List all investigations |
GET |
/api/v1/investigations/{id} |
Get investigation by ID |
DELETE |
/api/v1/investigations/{id} |
Delete investigation + all its signals |
GET |
/api/v1/investigations/{id}/stream |
SSE real-time event stream |
DELETE |
/api/v1/investigations/failed |
Bulk-delete all failed investigations |
GET |
/api/v1/watchlists |
AML watchlist patterns |
GET |
/api/v1/signals |
All risk signals across investigations |
GET |
/api/v1/entities |
Entity profiles |
POST |
/api/v1/entities/{id}/notes |
Add analyst note to entity |
GOOGLE_API_KEY= # Google AI Studio key (paid tier required for Search grounding)
MONGODB_URI= # Atlas connection string (mongodb+srv://...)
MONGODB_DATABASE=hawkai
GEMINI_MODEL=gemini-2.5-flash
SYNTHESIS_MODEL=gemini-2.5-pro
EMBEDDING_MODEL=gemini-embedding-001
DISABLE_GOOGLE_SEARCH=falseMIT