simpliLink: AI-Native ERP Integration Middleware for the Modern Manufacturing Stack

Enterprise ERP integration is a solved problem — at least, that’s what your SI vendor will tell you. The reality inside tier-2 and tier-3 manufacturing plants is very different: brittle point-to-point connectors, hand-rolled ABAP scripts, and decades-old middleware that breaks every time a vendor pushes a schema update. The data exists. The systems exist. The gap between them is where operational efficiency dies.

simpliLink is Simplico’s answer: an AI-native ERP integration middleware built on a modern async stack, with a RAG-powered intelligence layer sitting directly at the integration seam.

This post is a technical deep-dive into the architecture, design decisions, and the AI capabilities that differentiate simpliLink from conventional integration platforms.

The Problem with Conventional ERP Integration

Most ERP integration projects deliver one of two outcomes:

Point-to-point connectors — fast to ship, impossible to maintain. Every new ERP version, every schema change, breaks the connector. You end up with a spaghetti graph of n×(n-1) integrations.
Monolithic ESB/iPaaS — expensive, over-engineered, and operated by a team of specialists. MuleSoft, Dell Boomi, SAP Integration Suite: all require significant licensing and configuration overhead before a single byte moves.

Neither approach handles the semantic layer problem: two ERPs can both have a vendor_code field that means completely different things, mapped to completely different master data. No amount of ETL fixes that without a human in the loop — or until now, an AI.

simpliLink Architecture

simpliLink is built around three core layers:

flowchart TD
    A["ERP Source Systems (SAP / Douzone / Ecount / Legacy)"] --> B["Connector Layer (FastAPI Adapters)"]
    B --> C["simpliLink Core Middleware"]
    C --> D["AI Intelligence Layer"]
    C --> E["Message Queue (Redis / Celery)"]
    C --> F["Normalized Data Store (PostgreSQL + pgvector)"]
    D --> G["RAG Engine (Embeddings + Vector Search)"]
    D --> H["Schema Reconciliation Engine"]
    D --> I["Anomaly Detector"]
    F --> J["Target ERP / Downstream Systems"]
    E --> J
    G --> F

Layer 1: Connector Layer

Each ERP source has a dedicated FastAPI adapter. Adapters are stateless, containerized, and expose a unified internal event schema regardless of the source system. Current adapters: SAP RFC/BAPI, Douzone Bizon REST, Ecount API, flat-file (CSV/Excel) for legacy systems.

Each adapter implements three contracts:

fetch(resource, filters) — pull-based data retrieval
stream(resource, callback) — webhook/polling event stream
push(resource, payload) — write-back to source system

This uniformity means the core middleware never talks to a specific ERP — it only talks to adapters.

Layer 2: Core Middleware

The middleware is a FastAPI application with async task execution via Celery and Redis. Its responsibilities:

Routing — event fan-out to one or many target adapters
Transformation — field mapping, type coercion, unit normalization
Idempotency — every event carries a deterministic event_id; duplicate delivery is handled at the database layer with upsert semantics
Retry & DLQ — failed events go to a dead-letter queue with exponential backoff; alerting via PagerDuty webhook
Audit log — every transformation written to an immutable append-only audit table

flowchart TD
    A["Inbound Event"] --> B["Idempotency Check"]
    B -->|"Duplicate"| C["ACK + Discard"]
    B -->|"New"| D["Schema Validator"]
    D -->|"Invalid"| E["Dead Letter Queue"]
    D -->|"Valid"| F["AI Schema Reconciler"]
    F --> G["Field Mapper"]
    G --> H["Transformation Engine"]
    H --> I["Outbound Adapter"]
    I -->|"Success"| J["Audit Log"]
    I -->|"Failure"| E
    E --> K["Retry Worker"]
    K --> F

Layer 3: AI Intelligence Layer

This is where simpliLink diverges from every conventional middleware on the market.

The AI Layer: Four Capabilities

1. Schema Reconciliation Engine

When you connect two ERPs built by different vendors for different markets, schema conflicts are guaranteed. simpliLink’s reconciliation engine uses an LLM to suggest field mappings based on field names, sample data, and domain context.

How it works:

At onboarding, each adapter’s schema is ingested and embedded into pgvector
When a new source field arrives with no explicit mapping, the engine performs a nearest-neighbor search across the target schema embeddings
The top-k candidates are scored with an LLM against sample values and field descriptions
Confirmed mappings are stored and reused; ambiguous ones are flagged for human review via the admin UI

This reduces initial mapping configuration from weeks to hours. The model learns your specific ERP vocabulary over time.

2. Natural Language Query Interface

Operations staff shouldn’t need to know SQL or navigate ERP sub-menus to answer basic operational questions. simpliLink exposes a chat interface backed by a RAG pipeline over the normalized data store.

flowchart TD
    A["User Natural Language Query"] --> B["Query Embedding"]
    B --> C["pgvector Similarity Search"]
    C --> D["Context Retrieval (Top-K Documents)"]
    D --> E["LLM Prompt Assembly"]
    E --> F["Answer Generation"]
    F --> G["Response with Source Citations"]
    G --> H["User"]

Example queries the system handles today:

"Which vendors had more than 3 late deliveries in Q1?"
"What’s our average GRN processing time for Korean-origin POs?"
"Show me all open work orders where BOM material is below reorder point."

All answers are grounded in the normalized data store — no hallucination risk on factual operational data, because retrieval is always the source of truth.

3. Anomaly Detection

simpliLink runs a lightweight anomaly detection layer over the event stream in near-real-time. Current detection rules (hybrid: rule-based + statistical):

Signal	Detection Method
Duplicate invoice	Exact-match hash + fuzzy amount/vendor comparison
Vendor price drift	Z-score over rolling 90-day average
Quantity mismatch (PO vs GRN)	Threshold with configurable tolerance %
Unusual approval timing	Time-series deviation from user baseline
Missing mandatory fields	Schema completeness check at ingest

Anomalies surface as prioritized alerts in the ops dashboard, with the AI layer providing a natural-language explanation of why the event was flagged and what the historical baseline looks like.

4. simpliDoc Integration (Document Intelligence at the Seam)

Purchase orders, delivery notes, quality certificates, and customs documents flow between ERP systems as PDFs or scanned images. simpliLink embeds simpliDoc — Simplico’s multilingual RAG document intelligence platform — directly in the ingest pipeline.

When a document arrives:

simpliDoc extracts structured fields (vendor, amount, line items, dates) via OCR + LLM parsing
Extracted fields are mapped to the normalized schema via the Schema Reconciliation Engine
The structured record enters the middleware pipeline — no manual re-keying

This closes the last mile of integration: unstructured documents that no conventional middleware can handle.

Deployment Architecture

simpliLink is delivered as a Docker Compose stack for on-premise or private cloud deployment — a deliberate choice for manufacturing clients with strict data residency requirements.

flowchart TD
    A["Client Network Perimeter"] --> B["simpliLink Stack (Docker Compose)"]
    B --> C["FastAPI Core (simplilink-core)"]
    B --> D["Celery Workers (simplilink-worker)"]
    B --> E["Redis (Queue + Cache)"]
    B --> F["PostgreSQL + pgvector"]
    B --> G["AI Service (Ollama / Claude API)"]
    C --> H["ERP Adapter Containers"]
    G --> I["LLM: Claude Sonnet (Cloud) or Llama3 (On-Prem)"]

The AI service is configurable: Claude API for cloud-connected deployments, Ollama with a local Llama 3 model for air-gapped environments. The middleware itself is LLM-agnostic — it calls a unified AI interface that routes to whichever backend is configured.

Performance Characteristics

Metric	Target
Event throughput	5,000 events/min per worker node
Schema reconciliation latency	< 3 seconds (LLM call)
NLQ response time	< 5 seconds (RAG + LLM)
Anomaly detection lag	< 30 seconds from event ingest
Data residency	Full on-premise option

Compliance & Security

simpliLink is designed with compliance-first principles relevant to manufacturing clients in Northeast and Southeast Asia:

Korea PIPA (개인정보보호법) — PII field tagging at the schema level; configurable data masking in the normalized store
Thailand PDPA — data subject access and deletion hooks in the audit log
ISO 27001 alignment — immutable audit trail, role-based access control, encrypted transport (TLS 1.3) and at-rest encryption
ISMS-P readiness — audit log export in formats compatible with Korean ISMS-P evidence requirements

What’s Next

The simpliLink roadmap includes:

Multi-tenant SaaS mode — for SI partners who want to resell simpliLink as a managed service
Pre-built Korean ERP connectors — Douzone Bizon iCUBE, Ecount ERP, and UFIDA (for China-Korea supply chains)
Active reconciliation — AI-suggested write-backs when discrepancies are detected, not just alerts
IATF 16949 quality data module — for automotive tier-2/3 supplier compliance in Korea and ASEAN

Conclusion

ERP integration is infrastructure. It should be invisible, reliable, and intelligent enough to handle the messy reality of enterprise data without a team of specialists babysitting it. simpliLink combines a production-grade async middleware core with an AI layer that learns your schema, speaks your language, and catches problems before they become audits.

If you’re a manufacturer running multiple ERP systems — or an SI looking for a middleware partner with a genuine AI differentiator — we’d like to talk.

Contact: tum@simplico.net | simplico.net

Simplico Co., Ltd. is a Bangkok-based software engineering studio with 10+ years of enterprise delivery experience across Thailand, Japan, and global markets. We specialize in AI/RAG applications, ERP integrations, cybersecurity tooling, and ecommerce platforms.