Skip to main content
Glama

šŸ›”ļø Edge-Cloud AIoT Sentinel (aiot-mcp-agent)

Author: Benalia Meriem (mi.benalia@gmail.com)
Specialization: Edge AI, IoT Systems, and Large Language Models
Academic Report: report/aiot_sentinel_report.tex | šŸ“„ Download Compiled PDF

Hierarchical Edge-Cloud AIoT Architecture with FastMCP, Multi-Modal Telemetry Ingestion, and Gated SLM Reasoning.
Built as a thesis-grade and portfolio-grade engineering project demonstrating modern AIoT, edge computing trade-offs, and deterministic agent safety.


šŸ“Œ Executive Summary & Research Questions

Deploying Large Language Models directly on raw IoT sensor streams creates two catastrophic engineering failures:

  1. Network & Token Flooding: Streaming high-frequency sensor floats (temperature, humidity, CO2, dB) consumes massive bandwidth and cloud token budgets on static, nominal data.

  2. Hallucinatory Actuator Hazards: Unchecked LLMs can dispatch dangerous commands to physical hardware (e.g., shutting down HVAC during thermal runaway, or disabling ventilation during chemical gas leaks).

šŸŽÆ Primary Research Questions (RQs)

  • RQ1 (Data Efficiency & Telemetry Gating): To what degree can on-device event-triggered deadband and rate-of-rise gradient filtering suppress redundant high-frequency IoT telemetry and LLM token overhead without attenuating critical transient anomaly transitions?

  • RQ2 (Deterministic Actuation Safety): Can a standardized protocol gateway (Model Context Protocol) enforcing pre-actuation deterministic interlocks eliminate safety invariant violations when probabilistic language models control physical actuators?

  • RQ3 (Latency & Multi-Modal Reasoning Trade-off): How does a tiered architecture combining edge event filtering with compact local Small Language Models (SLMs) perform relative to static heuristic baselines and monolithic cloud streaming in decision latency, cross-sensor false alarm rejection, and resource efficiency?

AIoT Sentinel solves these challenges with a three-tier hierarchical architecture:

  • Tier 1 (Physical Edge / TinyML Gate): Lightweight deadband & finite-difference rate-of-rise gradient filtering that eliminates 90.0% of redundant telemetry while deterministically catching critical edge transitions.

  • Tier 2 (FastMCP Gateway & Safety Interlocks): A Model Context Protocol (MCP) server providing typed tools, tamper-evident SHA-256 chained audit trails, and deterministic invariant checks (ISO-50001 building codes).

  • Tier 3 (Gated SLM Agent): Correlates multi-modal sensor signals (thermal, acoustic, gas, optical, contact) to discern genuine threats from false alarms.


Related MCP server: evav-gateway

šŸ›ļø System Architecture

System Architecture

[ ESP32 Sensor Fleet ] (DHT22, MQ-135, PIR, Acoustic, Reed Contact)
          │
          ā–¼  (Raw Telemetry Stream)
ā”Œā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”
│ Tier 1: TinyML Edge Filter                             │
│ • Deadband thresholding (Δx ≥ δ)        │
│ • Rate-of-rise gradient detection (dx/dt ≥ γ) │
│ • Periodic heartbeat synchronization                  │
│ ──> Discards 90.0% redundant nominal samples           │
ā””ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”˜
          │ (Escalated Event Packets Only)
          ā–¼
ā”Œā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”
│ Tier 2: FastMCP Gateway (Model Context Protocol)       │
│ • get_room_state(room_id)                              │
│ • query_safety_policy(domain)                          │
│ • execute_actuator_action(...)                         │
│   ā”œā”€ā”€> [ Deterministic Safety Interlock Gate ]         │
│   └──> [ Tamper-evident SHA-256 Chained Audit Trail ]  │
ā””ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”˜
          ā–²
          │ Tool Calls & Safety Justifications
          ā–¼
ā”Œā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”
│ Tier 3: Multi-Modal SLM Reasoner (Qwen2.5 / Edge SLM)  │
│ • Cross-sensor consensus reasoning                     │
│ • False alarm rejection (e.g. transient solar heating) │
│ • Multi-stage disaster mitigation planning             │
ā””ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”€ā”˜

šŸ“Š Empirical Head-to-Head Benchmark

Evaluated over 25 continuous temporal scenarios (1,500 total ticks / 9,000 multi-sensor observations) including thermal runaway, chemical gas leaks, after-hours intrusion, HVAC energy waste, and transient solar glare:

Architecture

Decision Accuracy

Safety Invariant Violations

Packets Transmitted

Bandwidth Saved

Avg Decision Latency

Total Tokens

Static Heuristics (Rule-based)

100.0%

2

9,000

0.0%

3.75 ms

0

Unfiltered Cloud LLM (Raw Stream)

92.0%

1

9,000

0.0%

485.00 ms

63,000

Hierarchical AIoT Sentinel (Ours)

100.0%

0

900

90.0%

21.11 ms

680

Key Empirical Findings:

  • Bandwidth & Token Efficiency: The edge filtering layer reduces transmitted telemetry packets by 90.0% and slashes token consumption by 92Ɨ (from 63,000 down to 680 tokens) with zero missed emergency anomalies.

  • Deterministic Zero-Violation Safety: While naive LLMs occasionally dispatch hazardous commands due to phrasing nuance, the FastMCP safety gate achieved 0 safety violations.

  • Edge Latency: Local SLM actuation completed in 21.11 ms, approximately 23Ɨ faster than round-trip cloud LLM streaming.


šŸ“ˆ Visual Benchmark Reports

Multi-Modal Telemetry & Edge Detection

Telemetry Stream

TinyML Bandwidth Reduction by Sensor Modality

Compression Breakdown

Head-to-Head Metrics Comparison

Benchmark Metrics

Latency vs Token Cost Trade-off (Pareto Frontier)

Pareto Frontier


šŸ› ļø Repository Layout

aiot-mcp-agent/
ā”œā”€ā”€ src/aiot/
│   ā”œā”€ā”€ simulator/
│   │   └── virtual_fleet.py       # Multi-room ESP32 telemetry simulator & AR(1) drift
│   ā”œā”€ā”€ edge_filter/
│   │   └── tiny_filter.py         # TinyML deadband filter & rate-of-rise gradient gate
│   ā”œā”€ā”€ mcp/
│   │   └── server.py              # FastMCP Server + JSON-RPC 2.0 + SHA-256 audit log
│   ā”œā”€ā”€ agent/
│   │   └── reasoner.py            # Multi-modal SLM correlation & decision agent
│   └── benchmark/
│       └── eval_harness.py        # 25-scenario continuous temporal evaluation suite
ā”œā”€ā”€ tests/                         # Full Pytest test suite (11/11 passing in 0.39s)
│   ā”œā”€ā”€ test_simulator.py
│   ā”œā”€ā”€ test_edge_filter.py
│   ā”œā”€ā”€ test_mcp_server.py
│   ā”œā”€ā”€ test_agent.py
│   └── test_benchmark.py
ā”œā”€ā”€ report/
│   ā”œā”€ā”€ aiot_sentinel_report.tex   # Formal academic IEEE-style paper (Author: Benalia Meriem)
│   ā”œā”€ā”€ aiot_sentinel_report.pdf   # Compiled publication-ready PDF report
│   └── figures/                   # High-res publication charts (300 DPI)
ā”œā”€ā”€ results/                       # Benchmark outputs (JSON and Markdown)
ā”œā”€ā”€ dashboard.py                   # Interactive Streamlit telemetry & live agent console
ā”œā”€ā”€ pyproject.toml
└── requirements.txt

šŸš€ Quickstart & Reproducibility

1. Installation

git clone https://github.com/Anish-Guntreddi/aiot-mcp-agent.git
cd aiot-mcp-agent
pip install -r requirements.txt
pip install -e .

2. Run Test Suite

pytest tests/ -v

3. Run Empirical Benchmark

python -m aiot.benchmark.eval_harness

4. Launch Live Streamlit Console

streamlit run dashboard.py

šŸ”¬ Hardware Deployment Guide (ESP32-S3 / MicroPython)

To bind this software gateway to physical hardware, flash the following snippet onto an ESP32-S3 board with a DHT22 and MQ-135 sensor:

# firmware/main.py (MicroPython for ESP32-S3)
import time, ujson, urequests
from machine import Pin, ADC

SERVER_URL = "http://<GATEWAY_IP>:8000/telemetry"
LAST_TEMP = 0.0

def read_and_filter():
    global LAST_TEMP
    # Read hardware sensor...
    current_temp = 22.4 
    if abs(current_temp - LAST_TEMP) >= 0.5 or current_temp > 40.0:
        payload = {"device_id": "esp32_01", "room_id": "server_room_a", "sensor": "temperature_c", "value": current_temp}
        urequests.post(SERVER_URL, json=payload)
        LAST_TEMP = current_temp

while True:
    read_and_filter()
    time.sleep(1)

šŸ“œ Citation & Academic Credit

If you use this work in academic research or engineering evaluations, please cite:

@article{benalia2026aiot,
  title={Hierarchical Edge-Cloud AIoT Architecture: Integrating TinyML Gating, Model Context Protocol (MCP), and Gated SLM Reasoning for Safety-Critical Facilities},
  author={Benalia, Meriem},
  journal={Department of Computer Science & Artificial Intelligence},
  year={2026}
}

šŸ“œ License

MIT License. Authored by Benalia Meriem (2026).

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      B
      quality
      B
      maintenance
      A governance proxy for AI tools — every MCP/agent tool call is policy-gated, secret-redacted, and written to a hash-chained, offline-verifiable audit trail.
      13
      MIT
    • A
      license
      Not graded
      quality
      D
      maintenance
      Governed MCP gateway that lets AI agents call tools with policy enforcement, prompt-injection screening, a kill-switch, and tamper-evident signed audit logs.
      Apache 2.0
    • A
      license
      Not graded
      quality
      B
      maintenance
      MCP server that lets AI agents interact with IP cameras via ONVIF/VAPIX, with fail-closed per-agent policy gates and hash-chained signed audit receipts.
      4
      MIT
    • A
      license
      Not graded
      quality
      C
      maintenance
      Enables AI agents to operate through a local, signed boundary that blocks prompt injection and secret leakage, verifies outputs, preserves cross-session memory, and provides offline-verifiable receipts. It also exposes 900+ MCP tools for discoverable agent actions.
      7
      MIT