Platform

Intelligence infrastructure for regulated credit

How B thinks, remembers, decides, and proves every action it takes.

The decision pipeline

1
Request
Borrower or lender action triggers an intent
2
Retrieve memory
Pull relevant context from 6 memory layers
3
Route to LLM
Select optimal model + Officer for the task
4
Get decision
AI generates structured response with confidence
5
Validate
Compliance guard checks boundaries + thresholds
6
Log
Merkle-chain audit trail, immutable and queryable

Average pipeline latency: 340ms for simple queries, 2.1s for full underwriting decisions. Every step is logged, timestamped, and attributable.

Memory architecture

Six layers of memory. Zero hallucination tolerance.

Working
Session

Active conversation context. What B is thinking about right now for this specific interaction.

Episodic
Permanent

Past interactions and outcomes. Every conversation B has had with this borrower or lender.

Semantic
Updated quarterly

Domain knowledge. Malaysian lending regulations, ANGKASA codes, product structures.

Procedural
Version-controlled

How to do things. Underwriting workflows, collection scripts, escalation protocols.

Meta
Real-time

Self-awareness. What B knows about its own capabilities, confidence levels, and limitations.

Social
Continuous

Relationship graphs. Who knows whom, organizational hierarchies, communication preferences.

Security and compliance

Built for regulated financial services

Data residency

All data stored in Malaysia. Primary infrastructure in MY-region cloud. No cross-border data transfer without explicit consent.

Encryption

AES-256 at rest, TLS 1.3 in transit. Per-client encryption keys. Hardware security module for key management.

PDPA compliance

Full Personal Data Protection Act compliance. Consent management, data subject access requests, right to erasure built in.

Merkle-chain audit

Every decision logged to an append-only merkle chain. Tamper-evident, time-stamped, queryable. Export for BNM reporting.

Portability

Provider-abstracted AI. No vendor lock-in.

B is model-agnostic. We route to the best available LLM for each task type and run an annual rotation drill to prove zero dependency on any single provider.

Multi-model routing

Each Officer task is routed to the optimal model. Underwriting may use a different provider than Collections. Cost and quality optimized per-request.

Provider abstraction

Standardized interface layer means switching providers requires zero client-side changes. Prompt templates adapt automatically.

Annual rotation drill

Once per year, we run B on an entirely different provider stack for 48 hours. Proves portability, benchmarks alternatives, keeps us honest.

Deploy Bolehlah.ai

Get B running on your loan book in under 2 weeks. We handle onboarding, data migration, and Officer calibration.