AI Engineer World's Fair 2025
Cognitive Shield Real Time Real Smart - Rachna Srivastava
Read the talk
Cognitive Shield Real Time Real Smart
A cloned voice can turn familiarity into an attack. Cognitive Shield connects regulatory workflows, specialized detectors, fraud graphs, and investigator tools to defend that trust.
From a talk by Rachna Srivastava
Before you start: Basic familiarity with databases, APIs, and machine-learning models is helpful; the article introduces the graph workflow and its query language.
When impersonation passes verification
Your manager calls. The voice has the familiar tone, urgency, and choice of words. The intranet is down, he says, and a client briefing cannot wait: send the confidential information to his personal email. Only after sending it do you discover that the caller was a voice clone. Rachna Srivastava opens with this scenario because the attacker succeeds by reproducing the cues an employee normally uses to establish trust.
The next example moves from conversation to identity verification. An AI-generated face appears during video KYC, blinks at the expected moment, smiles naturally, and clears identification. A verification process can become the entry point when its evidence is synthetic. These attacks do not necessarily break through a technical barrier; they satisfy the checks and enter as apparently legitimate users. Cognitive Shield is introduced as a defense against that combination of synthetic identity, convincing communication, and automated deception.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Three ways to manufacture trust
Srivastava tells three stories that move from an urgent phone call to a prolonged relationship and then a fabricated investment ecosystem. In her first account, Anthony, a retired father in California, receives a call from someone sounding unmistakably like his son. The caller describes an accident involving an injured pregnant person and says he needs bail. A second caller, posing as the son's lawyer, reinforces the emergency. Srivastava says Anthony wires $50,000, his entire retirement savings. She describes the voice as a clone made from his son's publicly available TikTok video; by the time the real son returns that evening, the money is gone. Familiarity establishes identity, and the supposed lawyer supplies a second layer of authority.
The second account concerns Lisa, who lives in Ohio and feels isolated after the pandemic. An Instagram contact claims to be a famous television star, calls her his soulmate, and promises marriage. Visa problems and financial difficulties explain why they never meet. Srivastava describes 18 months of daily messaging during which Lisa sends nearly $40,000. The face, she says, was AI-generated. She calls the scheme pig butchering: a relationship cultivated to extract money, with AI and cryptocurrency helping scammers conceal their tracks. Lisa reportedly described the experience in January 2025 to warn others. Here, repeated contact gradually substitutes for independent verification.
The third account follows Xavier, an accountant in Austin, who discovers Zipmax Pro in early 2025. Its credibility comes from an entire supporting environment: a polished website, enthusiastic YouTube testimonials, a white paper packed with AI and blockchain terminology, and an active Discord community. Weekly livestreams and AMAs feature synthetic avatars modeled on Silicon Valley influencers; a deepfake Elon Musk endorsement adds borrowed authority. The pitch promises up to 35% annual returns from AI-optimized DeFi investments.
Srivastava describes Xavier investing $60,000 of personal savings and his entire 401(k), without resolving whether those amounts overlap. The creators then dump their holdings in a rug pull, collapsing the coin's value. She says more than 5,000 people across the United States were defrauded. In this account, AI supports multiple parts of the operation: fake identity verification at exchanges, celebrity impersonation, plausible smart contracts, social bots, and synthetic influencers. The apparent corroboration comes from many surfaces, but those surfaces belong to the same manufactured story.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Using AI to recognize engineered deception
Srivastava reports a 375% increase in AI-powered scams since 2023 and says 76% of synthetic identities bypass traditional fraud detection. The presentation supplies no population or measurement method for those percentages. She also cites $9.3 billion in cryptocurrency losses reported to IC3 for 2024 and describes a 66% annual increase. The cryptocurrency total is broader than AI-enabled fraud; it should not be read as an AI-fraud subtotal.
The engineering motivation is the combination of emotional manipulation and scale. An attacker can generate credible identities and communications while reproducing the operation across many targets. The defensive premise is that models capable of learning behavioral patterns can also be trained to recognize manipulation. Cognitive Shield applies that premise across data management, detection, and response, rather than treating a convincing message or face as sufficient evidence of legitimacy.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
A defense spanning operations, detection, and response
Cognitive Shield organizes financial fraud defense into three layers. The first holds the operational data and assists users before problems become cases. The second examines activity for fraud. The third gives investigators the information and workflow needed to respond.
| Layer | Responsibility | Main capabilities |
|---|---|---|
| Secure operations | Establish and manage trusted records | Users, licensing, examinations, cases, payments, AI guidance |
| Fraud detection | Identify suspicious activity and connections | Specialized detectors, user and transaction graphs |
| Investigation and response | Turn signals into accountable action | Search, dashboards, escalation, audit trails |
The detection layer is described as eight modules covering threats such as deepfakes, bots, phishing, synthetic identities, and cryptocurrency scams. Graph analysis adds relationships among users and transactions. The response layer combines those signals with human investigation, including trend analysis, automatic escalation, and compliance records. Detection is one stage in a workflow that must also preserve evidence and assign responsibility.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
AI assistance inside regulatory work
The foundation layer stores licensing applications, examinations, case records, and payments in a secure database. AI assistance is attached to the work users already perform. When someone submits a license application or renewal, the system checks for missing information, flags inconsistencies, and offers guidance while the form is being completed. During examinations, it reviews responses and documents for unusual patterns that deserve human attention.
Legal and billing assistance explains case outcomes, fines, and deadlines in ordinary language. A built-in assistant accepts natural-language questions and uploaded legal documents for summaries and insights. Role-specific dashboards then expose application status, compliance status, and payment workflows to regulators, licensers, and auditors. The intended benefit is continuity: a user can move from a record to an explanation and then to the next required action without leaving the operational context.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Different detectors for different evidence
The detection layer uses specialized components because the evidence for a cloned voice differs from the evidence for coordinated blockchain activity. The module slide names Deepfake Hunter, Graph Intelligence, Pattern Evolution, Social Media Guardian, Communication Shield, Bot Detection, Blockchain Watchdog, and Payment Guardian. The spoken explanation groups capabilities by threat and model family rather than giving a separate implementation for every label.
| Detection task | Described technique | Evidence examined |
|---|---|---|
| Deepfakes | Generative adversarial network (GAN)-based systems | Manipulated media |
| Bots | Gradient boosting machine classifiers | Automated blockchain activity |
| Phishing | NLP plus WHOIS and other checks | Communication and domain information |
| Cryptocurrency scams | Graph neural networks (GNNs) | Transaction networks and anomalies |
The GAN reference identifies a model family without specifying the detector's training procedure or decision rule. For phishing, language analysis is combined with information about the domain; for cryptocurrency fraud, relationships among transactions become part of the evidence.
More broadly, deep learning analyzes images and audio for deepfakes and voice cloning. GNNs connect users, devices, and transactions to reveal possible fraud rings. NLP examines phishing, social engineering, and unusual communication patterns. Multimodal signal processing combines text, voice, and metadata so that an investigation can consider several kinds of evidence together. The presentation describes these capabilities, but gives no measured accuracy, false-positive rate, throughput, or end-to-end latency.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
From unstructured evidence to a queryable fraud graph
Fraud may involve a network of accounts, people, and devices rather than a single suspicious record. The graph workflow therefore begins with a data-construction problem: emails, PDFs, forms, and logs do not arrive as nodes and edges. Their entities and relationships must first be extracted and connected.
Step 1: Build and enrich the graph. CrewAI coordinates an agentic workflow using large language models to extract entities and relationships from text, PDFs, forms, emails, and logs. Internal Postgres records enrich that graph. Srivastava gives two examples of what models such as GNNs can then look for: accounts acting in sync and devices reused across multiple fake identities. The graph makes a shared device an explicit connection between accounts, rather than leaving an investigator to discover it across separate records.
Step 2: Persist the relationships. Neo4j stores the graph's nodes and edges. Step 3: Ask questions in natural language. An LLM translates an investigator's question into Cypher, the graph query language understood by Neo4j. Srivastava describes this as a Neo4j-based RAG system that surfaces patterns, anomalies, and entity linkages. Here, GraphRAG names the retrieval approach; the talk does not identify a particular GraphRAG package or version.
For a concrete illustration of the reused-device question, suppose accounts A101 and A102 both have a USED_DEVICE relationship to device D7. An investigator asks, “Which devices are shared by multiple accounts?” With illustrative Account, Device, and USED_DEVICE schema names, the corresponding query is:
cypher
MATCH (account:Account)-[:USED_DEVICE]->(device:Device)
WITH device, collect(DISTINCT account.id) AS account_ids
WHERE size(account_ids) > 1
RETURN device.id AS device_id, account_ids
ORDER BY device_id;
For those example records, the query would return D7 with A101 and A102. That is a relationship worth investigating, not proof that either account is fraudulent. The useful transition is from scattered account evidence to an explicit, queryable connection.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Turning alerts into accountable action
The response layer begins with a unified fraud intelligence console. Investigators use natural-language search to access information in Postgres and Neo4j without remembering each database's query syntax. Live dashboards show fraud hotspots, changing tactics, and connections among suspected bad actors, bringing the evidence into one investigative view.
Escalation adds a decision about ownership. A combination of rules and LLM-based logic assesses the severity of an open case and routes it to an appropriate person or team. Role-based access controls who can act, while an audit trail records what happened. Investigations can then produce PDF or CSV reports for regulators, auditors, and internal teams. The operational chain is signal, assessment, routing, investigation, and a traceable record of the response.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Walking through the operational application
The application walkthrough starts on a dashboard with recent searches, emerging threats, and recent alerts. The visible Recent Searches table includes result counts and risk-score bars. Navigation follows the same three-layer structure: user management, fraud and graph detection, then investigation, search, dashboards, and alerting.
The first group of screens covers administrative work:
- Users and cases: User activity, security settings, accounts, and contacts sit alongside case creation, AI case analysis, and case search.
- Examinations: Scheduled, in-progress, completed, and canceled examinations have management views, with risk assessment, risk history, and business rules. Timing, expenses, billing, analytics, and reporting accompany these workflows.
- Invoices: A summary dashboard separates draft, sent, paid, overdue, and canceled invoices. Available actions include creating an invoice, exporting it, and requesting AI analysis.
The repeated dashboard-first layout gives each workflow a summary before its individual records and actions.
Processing screens link organizations, perform calculations, and offer AI discrepancy detection. The payment portal exposes pending payments, payment history, and receipts. A multimodal assistant then adds voice input and output, web search, and the ability to browse and save user data for later search. These screens connect the earlier description of AI assistance to the records and statuses an operator actually handles.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Monitoring media, payments, and online activity
The fraud screens organize evidence by its source and the type of investigation:
- Deepfakes: Image, video, and audio analysis appear alongside recent analyses, search, and threat intelligence.
- Payments: Views cover suspicious transactions, active cases, response time, trends by fraud type, and alert statistics, with AI payment analysis and detected patterns.
- Web and social media: Web fraud has real-time monitoring. Social views let an operator choose platforms and examine investment, phishing, and cryptocurrency scams.
- Cryptocurrency: Screens expose suspicious wallets, smart contracts, transaction flows, blockchain activity, financial impact by blockchain, and fraud trends.
- Phishing and bots: Phishing views show active campaigns, domain analysis, and targeted organizations. Bot views expose networks, behavior, and threat analytics.
The walkthrough demonstrates the monitoring surfaces available to an investigator; the displayed categories and charts are not an evaluation of detector performance.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Graph construction and investigation tools
The knowledge graph builder connects to Neo4j and offers text-to-graph construction from user text, PDFs, and document files. A query generator accepts natural language and produces Cypher; a GraphRAG interface provides retrieval and insights. These are the application counterparts of the earlier build, persist, and query workflow.
Investigation and enforcement add general search, domain regulation monitoring, and search history. System analytics include AI insights about transactions and alerts, fraud patterns, geographic threat distribution, and risk-model performance. Alert monitoring filters by severity, status, and source, while retaining attack records and statistics and supporting active investigations. The walkthrough ends with APIs used by the frontend to extract and manage data, establishing the application's API-driven structure before the implementation stack is introduced.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
The implementation stack
The stack separates presentation, API handling, agent coordination, and storage:
| Component | Role in Cognitive Shield |
|---|---|
| Streamlit | Dashboard frontend |
| FastAPI | Incoming login activity, uploads, and other API data |
| CrewAI | Coordination of collaborating AI agents |
| Postgres | Operational data storage |
| Neo4j | Graph storage and analysis |
| GraphRAG and LangChain | Retrieval and agent functionality |
Srivastava presents this combination as the basis for scaling real-time detection. Operational records remain in Postgres while Neo4j supports relationship analysis; the graph complements the transactional data layer.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Making fraud decisions observable and explainable
The implementation lessons begin with security: trust must be designed into the system from the start. Model specialization follows. Fraud changes quickly and spans different evidence types, so Srivastava recommends collaborating agents trained for particular tasks rather than relying on one model to recognize every kind of attack. Graphs add a way to inspect hidden connections alongside relational records.
For service design, she recommends an API-driven architecture and FastAPI microservices instead of a large monolith. Operationally, the system must monitor model behavior as well as uptime, including false positives and false negatives. Decisions need to be tracked and explainable so investigators can understand why the system acted. Privacy belongs in the initial design too; Srivastava describes the team's approach as encrypting everything and assuming nothing. A fraud alert earns trust through inspectable evidence and decisions, not simply through the presence of AI.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Defending trust, and the teaching behind it
The closing returns to the full lifecycle: protect user data and assist people working with it; apply deep learning, graph AI, and LLMs to detection; then expose signals through response tools and compliance reporting. The graph recommendation is complementary: relational databases need graph support to capture networks of connections. Multi-agent LLMs are presented as a way to bring speed, clarity, and context to time-sensitive decisions.
Srivastava forecasts that by 2027, 90% of cyberattacks will be AI-driven and annual fraud losses will exceed $100 billion. These are forecasts offered to convey urgency, without a supporting methodology in the presentation. Her stated mission is prevention: intervene before fraud succeeds, and make defending trust a collective responsibility.
The final dedication explains the educational and ethical roots of that mission. Srivastava credits Jeremy Howard's classes and code at fast.ai with her deep-learning education, including the belief that AI should be open, ethical, and accessible. She names ULMFiT, the language-model fine-tuning work of Howard and Sebastian Ruder, as an influence on NLP. She connects Cognitive Shield to that teaching: practical model building joined to a responsibility for how the technology affects people.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Resources
From the talk
Current documentation for coordinating specialized agents through Crews and managing workflows with Flows.
Guides and API documentation for building interactive Python applications and dashboards.
Tutorials and reference material for building Python HTTP APIs.
Jeremy Howard and Sebastian Ruder's original ULMFiT paper on transferring pretrained language models to text-classification tasks.
Further reading
Examples of graph retrieval, including natural-language questions translated into Cypher and answers grounded in returned records.
Jeremy Howard's free practical course covering model training, deployment, and applications across images, text, and tabular data.
Full transcript
Loading transcript…