Redis Real-Time Decisioning Demo | Agentic Loan Onboarding

Stage 1: The Architecture

Agentic underwriting powered by Redis Context Retriever MCP tools

Credit bureaus, income verification, fraud signals, and policy data exist across separate systems. The AI underwriting agent does not query them directly. Redis Context Retriever generates MCP tools from the applicant schema — governed, structured interfaces the agent calls to assemble the Applicant 360 and reach a decision in under 15 milliseconds.

Data Sources

Credit Bureau APIs

Experian, Equifax, TransUnion — tradelines, scores, inquiries

Loan Origination System

Application data, requested terms, identity documents

Income Verification Platform

Payroll records, bank statement analysis, employer data

Fraud + Identity (KYC)

Identity confidence, device signals, behavioral biometrics

Policy Engine

Product eligibility rules, rate tables, concentration limits

Kafka / Event Stream

Application events, bureau pull triggers, workflow signals

Ingest Layer

RDI

Syncs bureau data, LOS state, and income signals into Redis

Redis FeatureForm

Serves applicant risk and eligibility features online

Unified Context Layer

Redis RAM

Hot application state, active bureau results, live session context

Redis Flex

Credit history, income records, prior application state, fraud embeddings

Redis FeatureForm

Feature store — credit quality, income stability, fraud risk, DTI capacity features

Redis Context Retriever

Generates MCP tools from the applicant schema: get_credit_profile(), get_income_signals(), get_fraud_risk(), get_policy_eligibility() — governed structured interfaces the AI underwriting agent calls directly. No custom integration code required.

Decision Engine

AI Underwriting Agent

Orchestrates MCP tool calls, interprets results, reaches decision

Risk Scoring Engine

Creditworthiness composite scoring and risk stratification

Policy Arbitration

Rate, term, product eligibility, and concentration limit rules

Redis Search

Peer segment matching and fraud typology similarity

Output Actions

Approve

Instant decision with rate and terms delivered in session

Step-up

Additional verification required before final decision

Counter-offer

Alternative product, amount, or term that fits risk profile

Decline

Adverse action with explanation and alternative guidance

Decision Target

<15ms

Primary Goal

Approval accuracy + conversion

Redis Role

Governed applicant context

Stage 2: Application Event

Sofia Reyes submits a $45,000 loan application — the AI agent begins calling Context Retriever MCP tools immediately

Sofia is a first-time Apex Lending applicant. She has submitted a digital application and is waiting on the confirmation screen. The AI underwriting agent has under 15 milliseconds to call Context Retriever MCP tools, assemble the Applicant 360, and return a decision before the confirmation screen renders.

Application Details

Sofia Reyes

First-time applicant | $45,000 personal loan | 60-month term

UNDER REVIEW

Eventloan_application_submitted

Requested amount$45,000 — 60-month term

Stated income$72,400/yr — Software engineer, 3yr tenure

Application channelDigital — mobile browser

Identity verificationKYC passed — identity confidence 0.94

Bureau consentTri-merge pull authorized

Policy pre-checkProduct eligible — personal loan, first-time applicant

Why This Moment Matters

Sofia is waiting on a confirmation screen. Every second without a decision increases the chance she abandons the application or accepts a competitor offer. The AI underwriting agent does not run a batch job — it calls Redis Context Retriever MCP tools in real time and reaches a decision while she waits.

The difference between this and a legacy underwriting stack is not speed alone. It is that the AI agent calls governed, schema-defined interfaces — not raw SQL against bureau data or custom API fan-outs. Context Retriever generates the tools from the applicant schema. The agent uses them.

Stage 3: Ingest

Bureau data, income signals, and fraud context flow into Redis

RDI synchronizes applicant state, bureau results, and income verification data. Redis FeatureForm serves the online features used by the AI underwriting agent. Context Retriever MCP tools are generated from the schema and ready to serve before the agent's first call.

Source Systems → Redis

BCR

Credit Bureau APIs

Tri-merge bureau pull, tradelines, scores, inquiries — RDI keeps results always-fresh

→

LOS

Loan Origination System

Application data, requested terms, identity documents

→

INC

Income Verification Platform

Payroll records, bank statement signals, employer confirmation

→

KYC

Fraud + Identity

Identity confidence, device fingerprint, behavioral biometrics

→

POL

Policy Engine

Eligibility rules, rate tables, DTI limits, product constraints

→

KFK

Kafka / Event Stream

Real-time application events, bureau pull completions, workflow signals

→

Pipeline Status

Bureau data syncAlways-fresh context — RDI

Income verificationSub-second sync

Context Retriever toolsGenerated from schema — ready

Redis FeatureForm parity100%

MCP tool latencySchema-compiled at startup

Decision modeInline — AI agent orchestrated

Stage 4: Context

The AI underwriting agent calls four Context Retriever MCP tools — no custom SQL, no API fan-outs

Context Retriever has generated MCP tools from the applicant schema. The AI agent calls them in sequence. Each returns governed, structured context. Together they form the Applicant 360.

Redis Context RetrieverRedis RAMRedis Flex

MCP Tool Calls — Context Retriever

MCP

get_credit_profile()

CREST Score: 718 | 5 tradelines | 0 derogatories | 2 hard inquiries (6mo)

→

MCP

get_income_signals()

$72,400/yr verified | Employer: 3yr tenure | Stable classification

→

MCP

get_fraud_risk()

Identity confidence: 0.94 | Device: confirmed | 0 fraud flags

→

MCP

get_policy_eligibility()

DTI: 28.4% (limit 43%) | Product eligible | Rate tier: Standard

→

Assembled Applicant 360

Credit quality718 CREST Score — 0 derogatory marks, stable tradeline history

Income verified$72,400/yr — payroll-confirmed, 3-year employer tenure

Monthly capacity~$1,450 available after existing obligations

Fraud risk levelLow — identity confidence 0.94, device confirmed

DTI ratio28.4% — well within 43% policy ceiling

Policy fitEligible — personal loan product, standard rate tier

Context signal: Redis Context Retriever generates these four MCP tools from the applicant schema definition. The AI underwriting agent calls get_credit_profile(), get_income_signals(), get_fraud_risk(), and get_policy_eligibility() as structured governed interfaces — not raw database queries, not custom API integrations. That is the Iris difference.

Stage 5: Feature Serving

Applicant risk and eligibility features hydrate in milliseconds

Redis FeatureForm serves the online features that feed the AI underwriting agent's risk scoring: credit quality, income stability, fraud risk, DTI capacity, approval prediction, and rate optimization.

credit_quality_score

Composite credit quality relative to product risk threshold

0.740.4 ms

income_stability_score

Income reliability and employment tenure signal

0.880.3 ms

fraud_risk_score

Composite fraud and identity risk (lower is better)

0.060.4 ms

dti_capacity_score

Remaining DTI capacity within policy ceiling

0.710.3 ms

approval_rate_predictor

Predicted approval likelihood given full context

0.820.5 ms

rate_optimization_score

Rate tier fit based on risk profile and segment

0.790.4 ms

Feature Serving Performance

Features Hydrated

P99 Lookup

2.3 ms

Train / Serve Parity

100%

Stage 6: Decision

Approve at 8.9% APR — AI agent reaches decision from MCP tool results

The AI underwriting agent has called four Context Retriever MCP tools, hydrated 94 features, and evaluated three action paths. The decision is grounded in the full Applicant 360 — not a single bureau score.

#1 Winning Action

APPROVE

$45,000 approved at 8.9% APR — 60-month term

Credit quality, income verification, fraud signals, and policy eligibility all clear

MCP tool results support a straight-through approval. DTI at 28.4%, credit score 718, fraud risk 0.06, income verified. The decision is delivered in session while Sofia waits.

Action confidence0.87

#2 Escalate

STEP-UP

Income document upload requested before final decision

Valid path if income_stability_score were below threshold

Not triggered here — payroll-confirmed income signal meets the documentation standard. Requesting additional documents would add friction without improving decision quality.

Action confidence0.71

Suppressed

DECLINE

Application does not meet credit policy

Suppressed — all four MCP tool results clear policy thresholds

Declining a creditworthy applicant is the outcome a legacy threshold-only system is more likely to produce. The Applicant 360 assembled by Context Retriever makes the right outcome clear.

Action confidence0.04

Stage 7: Business Impact

The value is a better lending decision delivered in the session, not via email in two days

Inline Applicant 360 assembly means faster approvals, fewer false positives, and straight-through processing for creditworthy applicants — without replacing the existing underwriting stack.

Operational Value

Decision latency<15ms inline — vs. 2–5 day industry average

Straight-through processing0% manual review for this application

False positive reductionContext-driven vs. score-only reduces over-decline

Conversion impactDecision in session — Sofia sees approval before she closes the tab

MCP tool calls4 governed interfaces — no custom SQL or API integrations

Schema governanceRow-level access controls enforced server-side by Context Retriever

Per-Decision Outcome

2–5 days

manual review
queue

→

<15ms

in-session
approval

Stage 8: Outcome

Same applicant. Same data. Decision in the session instead of the inbox.

Without Redis Iris, the underwriting agent waits for custom API calls and batch bureau joins. With Redis Iris, Context Retriever MCP tools return governed structured context in milliseconds and the AI agent delivers the decision before Sofia closes the browser.

Legacy Underwriting Path

Sofia Reyes

Application in review

StatusApplication queued for manual underwriting review

Decision timeline2–5 business days — email notification

Context usedSingle bureau score pull — no income or fraud cross-check

2–5 days

decision

manual

review

at risk

conversion

Redis Iris-Powered Underwriting

Sofia Reyes

Preferred — Context Retriever MCP tools assembled

Approved

$45,000 at 8.9% APR — 60 months

4 MCP tool calls. 94 features. Applicant 360 assembled. Decision delivered in session.

Why this wins

Credit 718 + income verified + fraud clear + policy eligible

Context Retriever governed all four data sources. No custom code. Schema-defined MCP tools.

87%

confidence

<15ms

decision

approved

in session

Stage 9: Architecture Recap

Redis Context Retriever makes the AI agent schema-driven, not code-driven

Bureau APIs, income systems, fraud platforms, and policy engines stay in place. Redis Context Retriever sits between those systems and the AI underwriting agent, generating MCP tools that return governed structured context — no custom SQL, no bespoke API integrations.

Data Sources

Credit Bureau APIs

Experian, Equifax, TransUnion — tradelines, scores, inquiries

Loan Origination System

Application data, requested terms, identity documents

Income Verification Platform

Payroll records, bank statement analysis, employer data

Fraud + Identity (KYC)

Identity confidence, device signals, behavioral biometrics

Policy Engine

Product eligibility rules, rate tables, concentration limits

Kafka / Event Stream

Application events, bureau pull triggers, workflow signals

Ingest Layer

RDI

Syncs bureau data, LOS state, and income signals into Redis

Redis FeatureForm

Serves applicant risk and eligibility features online

Unified Context Layer

Redis RAM

Hot application state, active bureau results, live session context

Redis Flex

Credit history, income records, prior application state, fraud embeddings

Redis FeatureForm

Feature store — credit quality, income stability, fraud risk, DTI capacity features

Redis Context Retriever

Decision Engine

AI Underwriting Agent

Orchestrates MCP tool calls, interprets results, reaches decision

Risk Scoring Engine

Creditworthiness composite scoring and risk stratification

Policy Arbitration

Rate, term, product eligibility, and concentration limit rules

Redis Search

Peer segment matching and fraud typology similarity

Output Actions

Approve

Instant decision with rate and terms delivered in session

Step-up

Additional verification required before final decision

Counter-offer

Alternative product, amount, or term that fits risk profile

Decline

Adverse action with explanation and alternative guidance

Decision Target

<15ms

Primary Goal

Approval accuracy + conversion

Redis Role

Governed applicant context