Lightweight Multi-Agent RAG Engine & Offline AI Architecture for African SMBs | Neevly Academy
# Neevly Applied AI Engine & Multi-Agent RAG
Lightweight, offline-first multi-agent orchestration and vector retrieval system designed for low-bandwidth African SME workflows and practical developer education.
---
## Overview
In Sub-Saharan Africa, reliable internet connectivity and high cloud API costs create significant bottlenecks for deploying production AI systems. Neevly addresses this with a tiered hybrid architecture:
- **Tier 1 (Edge & Local Execution)**: Uses lightweight open-weight models (such as Gemma 2B/7B) to perform intent routing, query parsing, and AST security validation locally on edge hardware with sub-100ms response times.
- **Tier 2 (Embedded Vector Store)**: In-memory and SQLite-backed Dense Cosine Similarity indexing over local domain documentation, eliminating external vector database latency and dependency.
- **Tier 3 (Cloud DeepMind Escalation)**: Routes complex multimodal reasoning and codebase-wide queries to Google Gemini 1.5/2.0 Flash when an active connection is available.
---
## Architecture Flow
```
Inbound Query
│
▼
[1. Guardrail & Security Agent] ───► (Rejects malicious SQL / Prompt Injections)
│ (Validation Passed)
▼
[2. Vector RAG Retrieval Agent] ───► (Cosine Similarity over Local SQLite Vectors)
│ (Top-K Chunks + Confidence Scores)
▼
[3. Synthesis & Action Agent] ───► (Citation-grounded response)
```
---
## Neevly Academy Curriculum Structure
Neevly Academy is an engineering initiative based in Kumba, Cameroon, focused on training developers through practical system design and real-world software deliverables:
1. **Student Registration System**: Desktop database interface with SQLite schema normalization, transaction management, and AST validation.
2. **Organization Attendance Logger**: High-throughput event logging tool with offline persistence and CSV/Excel stream exports.
3. **Pharmacy Inventory & Stock Manager**: FastAPI service integrating inventory threshold alerts with vector-assisted search.
4. **Auton …