AI-Powered Chat Assistant for Blockchain Fraud Analysis

Results at a Glance

• 50× faster blockchain investigations compared to manual workflows
• Zero SQL dependency for AML analysts and investigators through an AI-powered investigation platform
• Single source of truth for transactional and behavioral blockchain data
• Built-in safeguards against hallucinations using evaluation datasets
• Enterprise-ready security model with strict data access controls

Part I. Overview of the Project

Client Background

The client is a highly specialized financial risk and compliance firm operating on the front lines of anti-money-laundering and crypto-fraud investigations, where AI for financial crime detection is becoming a critical capability. Their analysts handle complex, high-stakes cases that involve tracing funds through dozens of wallets, interpreting token flows, understanding the timing and intent behind transfers, and documenting evidence for regulatory reporting.

Despite their deep expertise, their daily work was constrained by the limitations of their tools. Analysts relied on such public blockchain explorers as Etherscan and a handful of disconnected dashboards. Investigations required dozens of open tabs, manual note-taking, copying transaction hashes into spreadsheets, and hoping that no critical link was missed along the way. Every case felt like assembling a puzzle without being sure all the pieces were available. The analysts knew exactly what questions they needed to ask, but had no system capable of answering them. That was a problem.

The Challenge

The entire investigative pipeline was built on manual work. Public tools simply weren’t designed to handle questions that AML teams face every day, creating a clear need for AML investigation automation. Even seemingly straightforward queries like “Show me wallets that received substantial inflows and fragmented them across several addresses within a short block window” required hours of manual tracing. More advanced questions were effectively impossible without a data scientist.

The root of the problem was fragmentation. Critical data lived across public explorers, dashboards, exports, and spreadsheets. There was no unified, high-performance data model capable of supporting rapid, iterative investigation. Without this backbone, analysts were left with instincts but no way to quickly validate hypotheses. The firm needed more than a dashboard. They needed a platform capable of understanding investigative intent and responding in real time.

Project Goals

Our mandate was clear: create a unified blockchain intelligence platform that would let every analyst ask complex, domain-specific questions in plain English and receive transparent, validated SQL queries in response. The platform had to be fast, accurate, compliant, and capable of supporting the type of deep forensic work that previously required specialized engineering help.

The goals centered around three pillars:

1. A single source of truth for all blockchain activity relevant to investigations.
2. A conversational AI assistant capable of translating natural language into optimized SQL, clarifying intent when needed, and explaining its reasoning.
3. A fully auditable analytical workflow where results could be trusted, documented, reproduced, and reviewed for regulatory purposes.

Part II. Technology Solution. Building an AI Assistant for Blockchain Fraud Analysis

Building the Foundation: A Unified Blockchain Warehouse

The project began with solving the most fundamental blocker: fragmented data. We architected a Snowflake-based warehouse explicitly designed for blockchain forensics, applying Snowflake blockchain analytics as the system’s analytical backbone. Instead of treating raw blockchain data as a flat log of transactions, we built a semantic layer that captured behavioral patterns, priced net flows, decoded contract interactions, and established clear, queryable relationships across entities.

For the first time, analysts gained access to a structured environment in which transactions, logs, events, transfers, and wallet relationships were integrated into a cohesive analytical model. Everything from historical token valuations to multi-hop flow reconstruction was available instantly.

Developing the Conversational AI Assistant

Only after the data foundation was complete did we begin constructing the conversational interface that would live on top of it. The assistant was built using Python and LangChain, forming a LangChain AI assistant capable of multi-step reasoning over complex blockchain data. The goal was not merely to generate SQL, but to understand the logic behind each investigative question and translate that logic into a precise, performant query.

The core challenge was intent understanding. AML questions aren’t generic; they encode patterns, temporal relationships, token semantics, and investigator intuition. To address this, we trained the model on the client’s semantic schema and the investigative patterns that are most relevant to financial crime work. When an analyst asked about wallets “splitting inflows” or “rapid dispersion after large receipts,” the assistant knew these referred to particular, observable behaviors in blockchain data, not broad or ambiguous concepts.

Clarity Through Dialogue

Ambiguity is inherent in natural language. The assistant was designed to engage analysts when their questions needed clarification. Instead of making assumptions, it asked targeted follow-ups, such as whether to consider incoming or outgoing flows, which time windows to apply, or whether contract internal transactions should be included. This dialogue helped ensure that the SQL produced was both accurate and aligned with investigative intent.

Transparency, Validation, and Trust

One of the most important features of the system was transparency. Every response included:

• the generated SQL,
• a natural-language explanation of the reasoning behind it,
• the assumptions made,
• and references to the relevant tables used in the query.

To further reinforce trust, we built a comprehensive automated QA suite containing dozens of “golden questions” with validated SQL results. This suite ran as part of every update, ensuring stability and preventing regressions – a crucial requirement for regulated, audit-heavy workflows typical of an enterprise AI assistant for financial crime.

A High-Level Architecture Snapshot

The solution consisted of four tightly integrated layers:

1. Snowflake warehouse: High-performance, semantically rich blockchain data source. 
2. LangChain reasoning layer: Interprets analyst questions, orchestrates tool calls, generates and validates SQL logic, and manages reasoning flow. 
3. Python execution layer: Implements LangChain workflows, enforces security controls, executes queries, and handles system integrations.
4. Analyst-facing conversational interface: A clean, intuitive environment enabling real-time investigation.

This architecture provided the speed, scalability, and reliability required for mission-critical investigations.

Part III. Value Delivered

A Fundamental Shift in How Investigations Happen

The introduction of the AI assistant for blockchain investigations significantly improved the client’s investigative workflow. Tasks that previously required more than 48 hours of work from a data scientist and were effectively impossible for analysts without SQL skills can now be completed in under five minutes. Analysts can move from a natural-language question to a validated result within a single session, without leaving the assistant interface.

Empowering the Entire Team

Before the system’s deployment, only 3-4 data scientists could write the SQL required for complex, ad hoc investigations. After rollout, all 25 analysts on the fraud team gained the ability to perform sophisticated analysis independently through the conversational interface. This eliminated the long-standing bottleneck and allowed the data science group to reallocate time to building predictive models rather than handling routine query requests.

A Trustworthy, Compliant Analytical Process

The platform’s built-in transparency, including the ability to review generated SQL and read natural-language explanations of query logic, improved trust and usability across the investigative team. The automated QA suite, powered by a large set of validated “golden questions,” ensured consistent and reliable results. This combination enabled analysts and compliance managers to rely on the system for business-critical investigative and reporting tasks.

Impact Summary

With the custom AI assistant and unified data warehouse in place:

• Complex investigations became 50× faster,
• Analysts gained full autonomy in performing advanced data analysis,
• Identification of sophisticated money-laundering patterns increased due to rapid, iterative exploration,
• And the entire workflow became more auditable, consistent, and scalable.

The platform is now a core component of the client’s investigative operations, supporting reliable, repeatable, and real-time AML analysis at scale.

Need to Build an AI Assistant for Complex, High-Stakes Analytical Work?

AML investigations depend on the ability to think freely, follow leads, and ask complex questions without waiting days for technical support or piecing together incomplete information. The combination of a unified blockchain warehouse with a domain-aware conversational AI assistant helped the client unlock a new level of investigative capability, one defined by speed, depth, and confidence.If your company relies on analysts who work with scattered data sources or are constrained by slow, manual tooling, a purpose-built AI assistant can transform your operational efficiency. Broscorp provides custom AI assistant development services that combine technical rigor with real-world usability, delivering systems that allow your teams to ask better questions and get better answers.