Fintech Quickstart

Reference architecture for putting Interven in front of LangGraph agents in regulated financial workflows — SOC 2, PCI, GLBA-aligned.

This guide is for fintech engineering teams building production AI agents for back-office work — KYC review, transaction monitoring, customer support, loan underwriting, compliance summarization, internal account lookups — who need policy enforcement and audit evidence that survives a SOC 2 / PCI audit.

It assumes you've already integrated LangGraph or are about to.

What you're getting

Interven sits in the tool-call path of your LangGraph agents:

LangGraph agent
   ↓ (every tool call)
Interven scan API
   ↓ (decision: ALLOW / DENY / SANITIZE / REQUIRE_APPROVAL)
   → real Slack / Salesforce / Jira / internal API

Out of the box you get:

Inline policy enforcement — block agent calls that match deny rules, redact PII before it leaves your VPC, route money-movement actions to a human via Slack approve/deny buttons.
Audit trail of every decision — exportable to Splunk HEC, Datadog Logs, or flat CSV. SOC 2 evidence by design.
Per-API-key IP allowlist — production keys pinned to your VPC NAT range, rejected at the edge if used elsewhere.
Tenant data retention — automatic prune of decision history after 30 / 90 / 365 days per your written retention policy.
Read-only Auditor role — your compliance reviewer logs in, pulls evidence, cannot accidentally mutate state during the audit.

Reference architecture (production)

┌─────────────────────────────┐    ┌──────────────────────────┐
│  Your VPC                   │    │  Your SIEM               │
│                             │    │  (Splunk / Datadog /     │
│  ┌──────────────┐           │    │   self-hosted ELK)       │
│  │ LangGraph    │           │    │                          │
│  │ agent        │           │    └──────────▲───────────────┘
│  │ (your code)  │           │               │
│  └──────┬───────┘           │               │ Daily NDJSON export
│         │ HTTPS              │               │ via /v1/decisions/export
│         │ Bearer iv_live_*  │               │ (cron job, auditor RBAC role)
│         │                    │               │
│         ▼                    │               │
│  ┌──────────────┐  scan +   │  ┌──────────────────────────┐
│  │ Interven     ├───────────┼─►│ Interven SaaS            │
│  │ /v1/scan     │  decision │  │ api.intervensecurity.com │
│  └──────┬───────┘           │  │                          │
│         │ ALLOW / SANITIZE  │  │  - policy + risk engine  │
│         ▼                    │  │  - encrypted vault       │
│  ┌──────────────┐           │  │  - approval workflow     │
│  │ Real upstream│           │  │  - SIEM export           │
│  │ (Salesforce, │           │  └──────────────────────────┘
│  │  Slack, ...) │           │
│  └──────────────┘           │
└─────────────────────────────┘
                         │
                         │ REQUIRE_APPROVAL: Slack channel webhook
                         ▼
                  ┌──────────────────┐
                  │ #risk-approvals  │
                  │ Slack channel    │
                  │ (Approve / Deny  │
                  │  buttons)        │
                  └──────────────────┘

Step-by-step

# Find your VPC NAT egress IP — your agents call out from this CIDR.
# Example values; substitute your real range.
VPC_CIDR="203.0.113.0/24"

# In the Console (https://app.intervensecurity.com/api-keys):
#   - Click "Mint API Key"
#   - Name: "production-langgraph"
#   - IP allowlist: $VPC_CIDR
#   - Save the iv_live_... once.

If a leaked key gets used from outside the allowlisted range, the gateway returns HTTP 403 before even hitting the policy engine. The blocked attempt counter shows up next to the key on the API Keys page.

2. Wire LangGraph through Interven

Install:

pip install 'interven-langchain[langgraph,langsmith]'

In your agent code:

from langgraph.prebuilt import create_react_agent
from langchain_openai import ChatOpenAI
from langchain_core.messages import HumanMessage
from interven_langchain import InterventCallback

agent = create_react_agent(
    ChatOpenAI(model="gpt-4o", temperature=0),
    tools=[approve_disbursement, lookup_account, post_to_slack],
)

cb = InterventCallback(
    api_key=os.environ["INTERVEN_API_KEY"],
    on_block="raise",   # in fintech, prefer hard-stop on DENY
    runtime_type="langgraph",
)

result = agent.invoke(
    {"messages": [HumanMessage("Approve disbursement #4821 for $14,500.")]},
    config={"callbacks": [cb]},
)

Every tool call now scans through Interven first.

3. Set up policies for the most common fintech risks

These three policies cover ~80% of fintech back-office threat models. Add them via the Console (Policies → Create Policy → JSON tab):

a. Block any outbound message containing PII or secrets

{
  "name": "P1: Block PII/secret egress to messaging tools",
  "rules": [{
    "match": { "tool_name": "slack", "operations": ["post_message"] },
    "condition": { "has_data_class": ["PII", "SECRETS"] },
    "action": "DENY",
    "priority": 10
  }]
}

b. Require human approval for money movement above a threshold

{
  "name": "P2: Approval for high-value disbursements",
  "rules": [{
    "match": {
      "tool_name": "custom_proxy",
      "operations": ["create_disbursement", "transfer_funds", "approve_payout"]
    },
    "condition": {
      "body_contains_any": ["amount", "principal", "disbursement_amount"]
    },
    "action": "REQUIRE_APPROVAL",
    "priority": 20
  }]
}

The reviewer sees the request as a Slack card with Approve / Deny buttons. The agent pauses, the reviewer decides, the agent completes the task in the same conversation turn.

c. Mark INTERNAL content from Drive / Confluence as non-egressable

{
  "name": "P3: Block egress of internal docs flagged as confidential",
  "rules": [{
    "match": { "tool_name": "slack", "operations": ["post_message"] },
    "condition": { "has_data_class": ["INTERNAL"] },
    "action": "SANITIZE",
    "priority": 30
  }]
}

The INTERNAL classifier matches "internal only", "do not share", "confidential", "NDA required", "proprietary", etc. — the markers your finance team already puts on sensitive memos.

4. Connect Slack for approvals + alerts

Create a Slack app at https://api.slack.com/apps
Enable Incoming Webhooks → add to your #risk-approvals channel
Enable Interactivity & Shortcuts → set the Request URL to https://api.intervensecurity.com/v1/slack/interactions
Copy the Signing Secret from Basic Information
In the Interven Console (Alerts → Add Slack channel):
- Webhook URL: the incoming webhook from step 2
- Signing secret: from step 4
- Events: REQUIRE_APPROVAL + DENY + INCIDENT
- Min risk: HIGH (or any — your call)
Click Test — a test card lands in Slack with Approve/Deny buttons

When the agent triggers REQUIRE_APPROVAL, the on-call risk reviewer gets the card in Slack, clicks Approve, and the agent finishes the task. No context-switching to a separate dashboard.

5. Set retention to your written policy

Settings → Audit Log → Log Retention Period:

Choose	When
30 days	EU GDPR-strict tenants where the auditor wants minimum data
90 days	SOC 2 default minimum
365 days	PCI DSS Type 2 review windows
Unlimited	When your written policy is "indefinite, manual prune"

Save. The retention prune job runs daily and deletes events / completed approvals older than the window for your tenant. retention_last_pruned_at on the tenant row records when it last ran (visible to admin via Settings).

6. Set up the auditor user

When your SOC 2 / PCI auditor needs read-only access:

Settings → Team Management → Invite User
Email: their email
Role: Auditor (read-only, for SOC 2 reviewers)
Send them the temporary password

What the auditor sees:

✅ Activity, Decisions, Traces, Approvals (history), Incidents, Policies, Agents, Tools
✅ Audit Log section with Compliance evidence export card (date range + 5 formats)
❌ NO Tool Credentials, NO API Keys, NO Alerts in sidebar
❌ Every mutation endpoint returns 403

The auditor pulls evidence themselves; you don't have to manually export and email files.

7. Wire the SIEM export into your daily cron

If you'd rather pull decisions into your SIEM continuously instead of having the auditor pull manually:

#!/bin/bash
# /etc/cron.daily/interven-decisions-export.sh
# Runs as a service account with an Auditor-role API session cookie.

YESTERDAY=$(date -u -d 'yesterday' +%Y-%m-%dT00:00:00Z)
TODAY=$(date -u +%Y-%m-%dT00:00:00Z)

curl -sS -b "/etc/interven/auditor-session.cookie" \
  "https://app.intervensecurity.com/api/telemetry/decisions/export?format=splunk&since=${YESTERDAY}&until=${TODAY}" \
  | curl -sS -X POST "https://splunk.your-corp.internal:8088/services/collector/event" \
      -H "Authorization: Splunk ${SPLUNK_HEC_TOKEN}" \
      --data-binary @-

For Datadog, swap format=splunk → format=datadog and POST to your DD intake URL.

What you DON'T need to do (yet)

Self-host Interven — the SaaS at api.intervensecurity.com is multi-tenant with AES-256-GCM at rest, dedicated RBAC, and SOC 2-aligned controls. Self-hosted is supported via docker-compose but not required for most fintech buyers.
Custom IAM / SCIM — username/password and OIDC SSO (Google + Microsoft) are generally available; per-user roles included. SAML / SCIM provisioning is on the Enterprise roadmap — contact sales if it's a procurement blocker.

What to ask procurement / your CISO before scaling

These are the usual fintech-CISO due-diligence questions. Honest answers:

Question	Today
Where is decision data stored?	Hosted: UAE-region Postgres under our direct control. Self-hosted: 100% inside your infrastructure (single docker-compose).
Are webhook URLs / API tokens encrypted?	Yes. AES-256-GCM at rest, single-tenant key via `INTERVEN_CREDENTIAL_ENCRYPTION_KEY`.
Audit log immutability?	Append-only inserts; deletes only via the retention prune job (logged with `retention_last_pruned_at`).
SOC 2 attestation?	Type II in progress (target: Q4 2026). Letter of attestation available on request.
HIPAA BAA?	Available on Enterprise tier.
Right-to-deletion for EU tenants?	Yes — set retention to 30 days OR submit a DELETE request to privacy@intervensecurity.com.
SLA?	Free / Solo / Team: best-effort. Pro: 99.5%. Enterprise: contracted up to 99.9%.
Data residency control?	Hosted is UAE today; multi-region (EU, US) available on Pro and Enterprise. Self-hosted gives you full control.
SSO?	OIDC (Google + Microsoft) generally available. SAML / SCIM on the Enterprise roadmap — talk to us if it's a procurement blocker.

For complete compliance posture see Compliance. If anything is a blocker, email sales@intervensecurity.com — fintech design partners get priority on roadmap items.

On this page