Section E · Reference

Domain Context

The vocabulary and stakes for both target domains — fraud & identity verification and multimodal sensor AI. Enough to sound credible in interviews and ask good questions back.

A · Fraud & identity (fraud-domain-flavored)

Fraud/identity platforms sit between credit applications and lenders, scoring identity authenticity. Knowing the domain language is part of the loop.

Vocabulary

TermMeaning
Synthetic identityAn identity manufactured from real and fake components — e.g., a real SSN with a fabricated name and DOB.
First-party fraudA real person applies in their own name, intending to default. "Bust-out fraud" is the canonical pattern — pay down small balances, then max out and abscond.
Third-party fraudIdentity theft — someone else's data used to obtain credit.
Account takeover (ATO)Adversary gains access to a real customer's account.
KYCKnow Your Customer — regulated identity verification, US-driven by USA PATRIOT Act and CIP rules.
AMLAnti-Money Laundering — broader regulatory framework targeting transaction patterns indicating laundering.
SARSuspicious Activity Report — filing to FinCEN required when a bank detects suspected illegal activity.
BureauCredit bureau (Experian, Equifax, TransUnion) — historical credit data.
eCBSVElectronic Consent-Based SSN Verification — SSA service to confirm SSN/name/DOB match. (Some platforms were first to go live.)
OFACOffice of Foreign Assets Control — US sanctions enforcement; institutions must screen counterparties against the SDN list.
ChargebackA reversal of a card transaction, typically initiated by the issuer at the customer's request. A delayed fraud label.
Bust-outFraud pattern: build credit history, max out, vanish. Often a synthetic identity executed over months.
CIPCustomer Identification Program — required US bank policy specifying how new accounts verify identity.

Types of fraud relevant to identity-verification products

Synthetic identity fraud

Modeled by identity-verification platforms at the application layer. The hard problem: a synthetic identity can have real bureau data (because the SSN was issued and may be associated with light real history). The signals are usually structural: implausible age vs SSN issuance window, address-history sparsity, device or behavioral anomalies during application.

First-party fraud (intent)

Hardest to model because the applicant uses real, accurate data — only their intent is fraudulent. Signals are behavioral: bust-out patterns, geographic anomalies in transaction history, application-to-credit-pull timing.

Identity theft

Real victim's data used. Easier in some ways — the legitimate person and their patterns exist on record, the fraudster's behavior often diverges (different device, geography, application velocity).

Identity verification

The pipeline at a typical lender:

  1. Applicant fills out an application.
  2. Identity verification: name + SSN + DOB + address validated via bureau and eCBSV.
  3. Fraud scoring: fraud-domain-style score on the identity itself.
  4. Credit decisioning: combine identity confidence with credit history → approve / decline / manual review.
  5. If reviewed: a human investigator looks at the application, decides, often files documentation.

the fraud/identity score lives in step 3, with an output that influences whether step 4 trusts the input.

Regulatory context

Knowing these names lets you discuss the stakes credibly:

  • Bank Secrecy Act (BSA): foundational US AML law. Requires SARs, CTRs (Currency Transaction Reports), and CIP.
  • FFIEC: Federal Financial Institutions Examination Council — issues guidance banks follow.
  • FinCEN: Financial Crimes Enforcement Network — collects SARs and CTRs.
  • OCC, FDIC, Federal Reserve: bank regulators. Examine compliance programs periodically.
  • CFPB: Consumer Financial Protection Bureau — consumer protection in financial services. Fair-lending implications for AI models.
  • FCRA: Fair Credit Reporting Act — governs use of consumer reports for credit decisions. Models that use bureau data must comply.
  • ECOA / Reg B: Equal Credit Opportunity Act — prohibits discrimination in credit. Fair-lending model validation is a big deal.

A staff DS in fraud/identity doesn't need to be a compliance expert, but should know these names exist and how they shape model design (specifically: adverse-action notices, fair-lending considerations, model explainability for regulator review).

B · Multimodal sensor AI (Solutions-Engineering-flavored)

Multimodal-sensor AI platforms build foundation models for physical-world AI, fusing video, sensor traces, and time-series data. Customer base spans industrial monitoring, mobility, retail analytics, IoT applications.

Vocabulary

TermMeaning
MultimodalA model that processes multiple input types (image + text + sensor) jointly.
Sensor fusionCombining inputs from multiple sensors to produce a more reliable estimate than any single sensor.
Lens(Solutions-Engineering-specific) a configurable analytical operation on Newton — e.g., "count people," "detect anomalies in equipment vibration."
Edge inferenceModel runs on a device (camera, sensor, gateway) rather than the cloud.
Inertial sensors / IMUInertial Measurement Unit — accelerometer + gyroscope + sometimes magnetometer.
Computer vision tasksDetection, segmentation, tracking, action recognition, depth estimation.
Time-series classificationPredict a label from a window of time-series data (activity recognition, equipment failure).
Anomaly detectionFlag observations that don't fit the normal pattern — common ask in industrial monitoring.
SynchronizationAligning multiple sensor streams to a common time base.
Foundation model (multimodal)Large pretrained model handling multiple modalities — CLIP, GPT-4V, Gemini, etc. Multimodal-AI platforms position themselves in this category.

Use cases

Common applications for multimodal sensor AI platforms:

  • Industrial monitoring: predictive maintenance from vibration, temperature, acoustic sensors on machines.
  • Retail analytics: foot traffic, dwell time, queue lengths from cameras.
  • Mobility: driver behavior, fleet management, traffic flow analysis.
  • Safety: PPE compliance, fall detection, perimeter monitoring.
  • Process optimization: throughput analysis on assembly lines, anomalies in operations.
  • Healthcare: patient monitoring, gait analysis, sleep tracking.

What makes these hard

  • Customer data is heterogeneous — every customer's sensors and conditions are different.
  • Labels are scarce and expensive — manual annotation of video / sensor traces.
  • Edge cases are operationally critical (the rare event you must catch) but rare in data.
  • Customer expectations vary — some want point estimates, some want intervals, some want narratives.

Interview probes

Show probe 1: "What's synthetic identity fraud, and why is it hard to detect?"

An identity manufactured by combining real elements (often a real SSN, sometimes belonging to a child) with fabricated elements (name, DOB, contact info). Hard because synthetic identities can have legitimate bureau records — the fraudster builds credit slowly first. Detection signals are structural: implausible age vs SSN issuance window, address-history sparsity, application velocity at the same device or address, behavioral anomalies during application (typing patterns, copy-paste of fields).

Show probe 2: "What is eCBSV?"

Electronic Consent-Based SSN Verification — an SSA service banks can use, with the applicant's consent, to confirm SSN matches the name and DOB. Replaces older indirect methods.

Show probe 3: "Why is fair-lending validation a big deal for fraud models?"

Even though fraud models aren't credit-decision models, they affect who gets credit. If a fraud model declines or flags applicants from a protected class at disproportionate rates, that's a fair-lending issue (ECOA / Reg B). Staff DS work includes disparate-impact analysis, often comparing decline rates across racial and geographic groups, and identifying features that may proxy for protected attributes (ZIP code is the classic example).

Show probe 4: "What's a 'lens' at a multimodal-AI company, in your understanding?"

A configurable analytical operation on the multimodal-AI platform — basically a templated combination of prompts and parameters that performs a specific task ("count people in a region," "detect machinery anomalies," "classify activities"). The lens is what Solutions Engineers and customers actually invoke. The DS role configures lens parameters and the prompts driving them, per POC.

Show probe 5: "What makes multimodal sensor AI hard compared to a single-modality model?"

Three sources of difficulty. (1) Synchronization — aligning streams from sensors with different rates, clocks, and reliabilities. (2) Modality fusion strategy — how to combine an image with a vibration trace meaningfully. (3) Data heterogeneity — every deployment has different sensors, conditions, and labels, making transfer between deployments hard. The strongest practitioners reduce these to common scaffolding: standardized preprocessing, modality-agnostic feature extraction where possible, and prompt-based composition over a flexible foundation model.

✓ Check yourself

Answer from memory before opening each one.

  1. An applicant uses their own accurate name, SSN and address, pays small balances for months, then maxes out every card and disappears. What type of fraud is this, and why is it the hardest to model?

    Show answer

    First-party fraud, in the bust-out pattern. Every identity check passes because the data is real; only the intent is fraudulent. So the signals must be behavioural, such as transaction patterns, geographic anomalies and application-to-credit-pull timing. See types of fraud.

  2. Why can a synthetic identity pass a bureau check, and which structural signals still expose it?

    Show answer

    The SSN was issued and may carry a light, real credit history, and the fraudster builds more over time, so the bureau record looks legitimate. Look instead for an age that does not fit the SSN’s issuance window, sparse address history, device or application velocity and odd behaviour on the form. See types of fraud.

  3. Where in a lender’s pipeline does the fraud and identity score sit, and what does it influence?

    Show answer

    After the identity check on name, SSN, date of birth and address, and before credit decisioning. It does not decide the loan; it sets how far the credit step should trust the identity it was given. See identity verification.

  4. A fraud model declines applicants in some ZIP codes at twice the rate of others. It is not a credit model, so why does this still matter, and what do you do?

    Show answer

    Fraud decisions affect who gets credit, and disparate decline rates across protected classes raise fair-lending issues under ECOA and Regulation B. ZIP code is the classic proxy. Run a disparate-impact comparison of decline rates and find features that proxy for protected attributes. See regulatory context.

  5. A suspicious activity report names fraud, and a chargeback does too. How do they differ as training labels, and what do both imply for timing?

    Show answer

    A bank files the report with FinCEN after it detects suspected illegal activity, and a chargeback is a card reversal the issuer starts at the customer’s request. Both arrive well after the decision, so they are delayed labels and recent rows will look clean. See the vocabulary.

  6. A customer wants anomaly detection on pump vibration traces plus camera footage. Name the sources of difficulty you would scope before promising a result.

    Show answer

    Synchronising streams with different rates and clocks, deciding how to fuse an image with a vibration trace, and heterogeneity across the customer’s sensors and conditions. Labels are scarce and the critical event is rare. See use cases.

Takeaway

  • Synthetic identity, first-party and third-party fraud differ in what is fake, so they need structural, behavioural and anomaly signals respectively.
  • Bust-out fraud passes identity checks by design; intent shows only in behaviour over time.
  • The fraud score sits between identity verification and credit decisioning and sets how far the credit step trusts the identity.
  • BSA, FCRA, ECOA and Regulation B shape model design through adverse-action notices, explainability and fair-lending analysis.
  • A fraud model that is not a credit model can still create disparate impact, and features such as ZIP code can proxy for protected attributes.
  • Multimodal sensor work is hard because of synchronisation, fusion, heterogeneous deployments and scarce labels, so scope before committing.

With the vocabulary in hand, the next chapter is a timed drill across every topic in the guide. → Practice Interview Questions

Further reading