SentinelTrade: AI Compliance Terminal

SentinelTrade automates bank trade screening using Gemini to detect sanctions and dual-use goods in real-time, turning manual document checks into instant, audit-ready decisions.

Comment

Inspiration

The inspiration for SentinelTrade: AI Compliance Terminal stems from the critical, high-stakes nature of trade finance. Billions of dollars move through global shipping lanes daily, yet the compliance checks guarding these transactions often rely on antiquated document reviews. We saw an opportunity to bridge the gap between human expertise and the sheer volume of data, building a "Command Center" that provides defensible, grounded reasoning for every decision.

What it does

SentinelTrade is a specialized terminal that acts as a digital "second opinion" for bank officers. It ingests trade documents and performs three critical functions:

  • Dual-Use Detection: Identifies goods with potential military applications.
  • Live Sanctions Verification: Cross-references entities against live global databases (OFAC, UN, EU) using real-time grounding.
  • Logical Auditing: Uncovers fraud by spotting inconsistencies between invoices and bills of lading.

How we built it

SentinelTrade is a Zero-Latency Staging Environment designed for high-stakes banking. The architecture is split into three core pillars:

1. The Multi-Modal Brain

We leveraged Gemini 3 Flash to process structured data and unstructured telemetry. Using the googleSearch tool, the AI performs live grounding against active sanction lists.

2. The Logic Sandbox

To allow officers to test scenarios without contaminating the official audit log, we built a RAM-only sandbox. It uses a custom Risk Scoring Model defined by the following equation:

$$ R_{score} = \left( \sum_{i=1}^{n} w_i \cdot S_i \right) \times \Gamma_{geo} $$

Where:

  • \( w_i \) is the weight of a specific compliance category.
  • \( S_i \) is the severity assigned by the AI \( (0–100) \).
  • \( \Gamma_{geo} \) is a multiplier for high-risk geographic shipping lanes.

3. The Immutable Vault

Once a decision is verified, the record is hashed using SHA-256 and stored in the Vault. This ensures the audit trail is tamper-proof.

Challenges we ran into

  • Document Variability: Trade documents vary wildly. We fine-tuned system instructions to ensure the AI could handle everything from messy scanned PDFs to digital spreadsheets.
  • Hallucination Guardrails: To prevent the AI from "inventing" sanctions, we implemented a strict JSON Schema validation layer. If the output does not match our interface, the system triggers a manual override: $$ P(error) \rightarrow \text{Trigger Manual Audit} $$
  • Performance vs. Precision: Balancing the speed of the Flash model with deep reasoning required careful prompt engineering and temperature tuning.

Accomplishments that we're proud of

We are particularly proud of our "Thought Trace" feature. It provides the officer with a step-by-step logic path. Integrating real-time Google Search grounding ensures our AI is aware of sanctions updated only hours prior.

What we learned

  • Explainability is Everything: In finance, a "Black Box" is a liability. We used Gemini's internal "thought process" to generate Thought Traces.
  • The Power of Grounding: Search-integrated AI bridges the gap between static training data and real-time international law.

What's next for SentinelTrade: AI Compliance Terminal

  • [ ] Simulation Engine: Adjusting global policy weights to see portfolio-wide effects.
  • [ ] Deep Network Analysis: Uncovering Ultimate Beneficial Ownership (UBO) structures.
  • [ ] Predictive Port Analytics: Identifying high-risk transshipment patterns.

Built With

Share this project:

Updates