AI Predictive Risk
& Balance Sheet Intelligence

Preemptify AI is the institutional-grade 4-OS BFSI software suite — four standalone but interoperable Industry Operating Systems: Banking OS (Retail · Corporate · Treasury · Compliance), Lending OS (Origination → Underwriting → Collections → Restructuring), Insurance OS (Underwriting → 5-state Claims → Lapse Revival) and Wealth OS (Advisory · MF · PMS · AIF · Demat). Each module ships with its own AI co-pilots, Customer 360 with name-resolved Workbench, NBA recommender, AI risk assessment, multi-turn Ask-AI chat, custom product catalog, JSON rules engine and module-scoped RBAC.

Unlike descriptive CRMs, Preemptify is preemptive. It ingests live central-bank and treasury signals from 24+ sovereign authorities (RBI · BOT · MAS · ECB · Fed +), classifies them through the Preemptify Neural engine, and fuses the signal with proprietary balance-sheet math to surface preempted product recommendations — each one citing a numeric driver and the source headline. PII is redacted before every LLM call. No black box.

Four standalone Industry Operating Systems. Each is licensable independently, deploys with its own AI co-pilots, simulators and module-scoped RBAC — and all four interoperate on a single cryptographic audit spine, AsyncLocalStorage tenant isolation, the same Neural Signal Engine, and a unified PII-redaction boundary. Every module below follows the same template: Header → Metrics → Lifecycle → AI Capabilities → Featured Simulator → Visual Preview → Roles & Regulators.

4.1 Signal Preemption Pipeline (4 stages)

1. Ingestion — RSS pulled from 24+ central banks & treasuries (Fed, ECB, BoE, BoJ, RBI, PBoC, MAS, BOT, BOT, BSP, SBV, BCB, Banxico, SARB, SAMA, SNB, BoK, BoC, RBA, BoI, US Treasury, +).
2. Classification — Each headline scored on {policy_direction, magnitude, time_horizon, asset_impact[], affected_sectors[], confidence}. Deterministic JSON contract; no free-text drift.
3. Aggregation — Weighted preempt-vector across rates / FX / credit / equity / commodity axes. Formula: source_trust × confidence × magnitude × time_decay.
4. Recommendation — When pressure exceeds calibrated thresholds, a Preempted NBA card is generated and prepended to the entity’s recommendation queue with source headline + math rationale.

4.2 Model details

5.1 Authentication & Authorisation

• ·HMAC-SHA256 signed JWT-like session tokens; 12-hour TTL; rotation via re-auth.
• ·4 role tiers: master (platform-owner) · admin (compliance officer) · executive (relationship manager) · guest (read-only sandbox).
• ·Scope-based RBAC: retail · corporate · admin · master. Server enforces on every route handler.
• ·Login ledger-anchored: every successful + failed authentication writes a SHA-256 chained audit row.

5.2 Transport & perimeter

• ·TLS 1.3 enforced on production ingress; HSTS preload; X-Frame-Options DENY; CSP with no inline scripts (other than the JSON-LD block).
• ·Enterprise+ deploys air-gapped: outbound DNS allow-list only; LLM gateway can be routed through customer-owned proxy.
• ·Customer PII never leaves the perimeter. All identity fields are tokenised (P-XXXXX) before logging or LLM submission.

5.3 Tenant-boundary controls

• ·Every LLM call enters via the single isolation gateway — direct calls to the upstream model from route handlers are architecturally rejected (lint + code review).
• ·Cross-tenant requests (e.g. an Demo India session attempting to fetch a Demo Thailand Bank-held entity) return HTTP 403 with a tenant_boundary_violation code and a cross_tenant_attempt: true audit row.
• ·Per-tenant SHA-256 hash-chained audit (tenant_audit_ledger); independently verifiable via GET /api/admin/tenant-audit/verify.

• ·Tokenisation by default — Names, account numbers, SSN / PAN / NRIC are HMAC-tokenised before they ever touch a log, prompt or audit row. Reverse-lookup is held in a separate KMS-encrypted vault.
• ·Data residency — On Enterprise+, all storage (MongoDB, object store, logs) is bound to a customer-chosen region. India (Mumbai), Thailand (Cyberjaya), Singapore, Frankfurt, US-East all supported.
• ·Right to erasure — DPDP 2023 §8 · GDPR Art. 17 · CCPA §1798.105 all supported via a single POST /api/privacy/dsr handler. SLA: 30 days.
• ·Public Sandbox — synthetic data only. No real-customer PII is ever processed. The sandbox demonstrates the platform’s capabilities without exposing any institution’s book.
• ·Cookies — strictly-necessary by default. Analytics is opt-in and privacy-first (Plausible / Umami). No third-party advertising trackers.

The platform is engineered against the following frameworks. Each product template is regulator-tagged so authoring inherits the correct legal handle.

When the same LLM instance services multiple banking clients (e.g. Demo India Bank and Demo Thailand Bank), the system treats them as two entirely separate entities behind a soundproof wall. The AI never mixes regional regulations (RBI vs BOT rules), product criteria, or internal bank policies across tenants. This is enforced by three concurrent rules:

8.1 Verifiable evidence

• ·Concurrent self-test endpoint: POST /api/admin/tenant-audit/self-test — launches simultaneous IN/Demo India + TH/Demo Thailand Bank inferences and verifies tenant_tokens are distinct + correct.
• ·Per-tenant SHA-256 chain verifier: GET /api/admin/tenant-audit/verify?country=IN&bank=Demo India%20Bank.
• ·Compliance officer dashboard: /admin/tenant-isolation — runs the self-test live and renders the per-tenant audit chain.

9.1 Inbound — your core banking → Preemptify

• ·Batch CSV / fixed-width / Cobol export — daily SFTP drop into the Schema Aligner endpoint. Manifest hash-anchored.
• ·Real-time webhook — POST /api/ingest/balance-event for high-velocity events (new account, large transaction, term-deposit maturity).
• ·ISO 20022 / SWIFT MT — bi-directional translator built into the Multi-Currency Core.
• ·JSON delta feed — Kafka / Kinesis / Pub-Sub bridges available on Enterprise+.

9.2 Outbound — Preemptify → your stack

• ·NBA push — POST webhook into your CRM (Salesforce Financial Services Cloud, Microsoft Dynamics, custom). One row per recommendation with cited rationale.
• ·Email / SMS — optional SendGrid / Twilio / Vonage / customer SMTP. Templated, regulator-attribution-aware.
• ·PDF reports — board-grade markdown rendered to PDF for credit committee / risk committee meetings.
• ·Audit export — SHA-256 chain + per-tenant audit ledger streamable to WORM (S3 Object Lock, Azure Immutable Blob).

9.3 LLM gateway choice (Enterprise+)

• ·BYO Azure OpenAI · BYO AWS Bedrock · BYO GCP Vertex · self-hosted Llama-3.1 / Mistral / Qwen on customer GPU pool · air-gapped Anthropic Claude via customer-owned proxy.
• ·Switching is a single env var (LLM_GATEWAY_BASE) — no code change.

Standard 90-day pilot timeline. All milestones are atomic — if any gate fails, the deployment auto-rolls back.

11.1 Quick start (Sandbox)

11.2 On-Premise (Helm)

11.3 Environment variables

12.1 Tier comparison

12.2 Free + Enterprise sandbox isolation rules

Both tiers identically enforce the three-rule Clean-Slate Context Isolation. The Demo Environment Picker on /login binds each evaluator session to exactly one tenant scope:

• ·🇺🇸 USA · Preemptify Demo Bank — book filtered to US entities. NBA cards cite Fed SR 11-7 / FDIC / OCC.
• ·🇮🇳 India · Demo India Bank — book filtered to IN entities. NBA cards cite RBI Master Directions / IRDAI / SEBI.
• ·🇹🇭 Thailand · Demo Thailand Bank — book filtered to TH entities. NBA cards cite BOT / OIC / SEC Thailand.

Cross-tenant access during a sandbox session triggers HTTP 403 tenant_boundary_violation and writes a cross_tenant_attempt: true row to the audit ledger.

12.3 Public sandbox demo identities (limited access)

The public sandbox at preemtify.com is open to anyone — but every account is restricted to synthetic data inside a single tenant scope, with the three isolation rules enforced. The two published demo identities are sandbox-role only:

12.4 What the public sandbox can and cannot do