ERP-Native Analytical Intelligence Layer
Vision
To build an Industrial AI Platform that transforms ERP data into trusted, explainable, and proactive business intelligence.
Wirabumi does not aim to build a public chat-bot.
We aim to build:
A domain-bounded ERP Analytical Copilot that speaks natural language but thinks in structured business logic.
Mission
- Make ERP data conversational without sacrificing accuracy.
- Reduce dependency on external paid-token LLM APIs.
- Keep intelligence inside trusted enterprise boundaries.
- Combine deterministic computation with natural language explanation.
- Evolve from reactive reporting into proactive anomaly detection.
Core Architecture Philosophy
1. Determinism First
All numbers come from structured ERP computation.
The language model never invents business data.
2. LLM as Interface — Not the Brain
The model explains results.
It does not calculate, optimize, or decide.
3. Structured Data Warehouse as Source of Truth
ERP data from systems such as Openbravo is transformed into:
- Fact tables
- Dimension tables
- Aggregated KPI models
- Time-series comparison tables
The AI reads curated business metrics, not raw transactional tables.
4. Controlled Scope
The assistant only answers ERP-related questions.
Out-of-domain requests are intentionally rejected to preserve trust.
Technology Stack
Data Layer
- ERP Source: Openbravo (and compatible ERP systems)
- Structured Data Warehouse (PostgreSQL or equivalent)
- KPI Aggregation Models
- Time-series comparison engine
Business Logic Layer
- Deterministic SQL computation
- Variance analysis modules
- Optional rule engine integration (e.g., Drools)
- Baseline and threshold comparison logic
AI Layer
- Small self-hosted LLM (e.g., LLaMA 7B/8B or DeepSeek small variant)
- Context injection (structured prompt engineering)
- No large-scale retraining required
- Optional fallback to external API (if needed)
The LLM acts as a controlled narrative generator, not an autonomous decision engine.
Orchestration Layer
- Natural language intent detection
- Safe SQL generation
- Guardrails & validation rules
- Multi-tenant isolation
- Prompt governance & versioning
Phase 1 — ERP Analytical Assistant
Objective
Deliver a conversational interface for structured ERP analytics.
Capabilities
Users can ask:
- Why did sales drop in February?
- Which region contributed most to the decline?
- Compare this month with last year.
- What product category under-performed?
System Flow
- User submits a natural language question.
- System interprets intent.
- SQL query is generated against curated data warehouse.
- Deterministic engine computes results.
- Structured output is injected into LLM context.
- LLM generates an explanation narrative.
Example Scenario
User asks:
“Why did sales drop in February 2026?”
Computed Data:
- January Revenue: 12.4B
- February Revenue: 10.1B
- Bali region decline: 30%
- Product A decline: 40%
Assistant Response:
Sales declined by 18.5%, primarily driven by a 40% drop in Product A sales concentrated in the Bali region.
Key Characteristics
- No hallucinated numbers
- Fully auditable
- Repeatable results
- ERP-bounded intelligence
- Predictable infrastructure cost
This phase delivers:
ERP Copilot — Conversational Business Intelligence
Phase 2 — Anomaly & Fraud Detective Engine
Objective
Shift from reactive Q&A to proactive detection.
Expanded Capabilities
The system automatically detects:
- Abnormal sales variance beyond seasonal baseline
- Unusual discount patterns
- Suspicious transaction timing
- Margin anomalies
- Procurement outliers
- Inventory inconsistencies
Analytical Techniques
- Time-series baseline modeling
- Statistical variance thresholds
- Anomaly detection models
- Rule-based fraud indicators
- Pattern deviation scoring
Example Scenario
System detects:
- Historical February sales drop: 5–8%
- Current drop: 30%
Assistant reports:
An abnormal decline exceeding expected seasonal variance has been detected in the Bali region. The deviation is statistically significant and primarily linked to Product A.
Evolution of the AI Role
In Phase 2, the AI becomes:
- Risk narrator
- Executive summary generator
- Early warning assistant
But computation remains deterministic and statistical.
This phase delivers:
Industrial AI Monitor — Proactive Business Risk Detection
Strategic Positioning
Wirabumi Industrial AI is:
- Not a generic chat-bot
- Not AGI
- Not creativity-focused AI
It is:
ERP-Native Industrial Intelligence Platform
Deterministic at the core, Conversational at the surface.
Business Impact
For Customers
- Faster root-cause analysis
- Reduced dependency on BI specialists
- Early anomaly detection
- Executive-level insight in natural language
For Wirabumi
- Stronger SaaS value proposition
- Reduced token-cost dependency
- Enterprise-grade AI positioning
- Scalable multi-tenant intelligence architecture
GET IN TOUCH