Offline Agentic BMS — PoC

Multi-agent Battlefield Management System powered by Microsoft Agent Framework (MAF), running fully offline on a single-node Kubernetes cluster with local LLM inference via Ollama.

Architecture

Field Operator           Command Post (HQ)
      │                        │
      ▼                        ▼
┌─────────────────┐  ┌─────────────────┐
│ Walkie-Talkie │  │ BMS Dashboard   │
│ (voice I/O)   │  │ (cases + live)  │
└────────┬────────┘  └────────┬────────┘
         │ audio          │ SSE
         ▼                ▼
┌─────────────────┐  ┌─────────────────────────┐
│ Speech Service  │─▶│ BMS API (FastAPI)          │
│ STT + TTS (GPU) │  │ /voice, /cases, /stream   │
└─────────────────┘  └──────────┬──────────────┘
                               │
                               ▼
                ┌──────────────────────────────┐
                │ MAF Agent Platform (3 levels) │
                │ L3: Orchestrator ←→ CaseManager / FieldSpecialist │
                │ L2: FieldCoord ←→ ReconAgent / VehicleExpert      │
                │ L1: CameraAgent ∥ MeteoAgent (concurrent)        │
                └──────┬──────────┬──────────┬─┘
                       │ MCP      │ MCP      │ MCP
                       ▼          ▼          ▼
               ┌────────┐ ┌────────┐ ┌──────────┐
               │Camera   │ │Weather │ │ BMS      │  ← All FastMCP v3.1+
               │(mock)   │ │(mock)  │ │(Postgres)│
               └────────┘ └────────┘ └──────────┘

All external system access from agents goes through MCP (Model Context Protocol). MCP servers built with FastMCP ≥ 3.1.

Tech Stack

Component	Technology
AI Framework	Microsoft Agent Framework (MAF)
LLM	Ollama + qwen2.5:7b (GPU)
MCP Servers	FastMCP ≥ 3.1
STT	faster-whisper (GPU, Spanish es-ES)
TTS	Piper TTS (GPU, Spanish es-ES voices)
API	FastAPI
Database	PostgreSQL
Frontend	HTML + SSE (walkie-talkie + dashboard)
Observability	Prometheus + Grafana Tempo + Grafana
K8s	kubeadm v1.35, single-node, GPU time-slicing
GitOps	ArgoCD + Gitea + Nexus

Project Structure

├── docs/                    # PoC specification and implementation plan
├── src/                     # Agent core (MAF workflows + agents)
│   ├── agents/              # 6 agent factories
│   ├── workflows/           # 3-level nested workflow builders
│   ├── config.py            # Environment-driven settings
│   ├── client.py            # Ollama client factory
│   ├── patch_ollama.py      # MAF bug #4402 workaround
│   └── runner.py            # CLI entry point (interactive + demo)
├── mcp_services/            # FastMCP servers (Camera, Weather, BMS)
├── bms_api/                 # BMS REST API (FastAPI)
├── speech_service/          # STT + TTS service
├── frontend/                # Web interfaces
│   ├── walkie_talkie/       # Field operator — push-to-talk voice UI
│   └── bms_dashboard/       # Command post — cases + live timeline
├── k8s/                     # Kubernetes manifests
│   ├── shared-services/     # Ollama deployment
│   ├── bms-ops/             # BMS application resources
│   └── monitoring/          # Tempo, ServiceMonitors, dashboards
├── prototypes/              # MAF pattern validation scripts (reference)
├── Dockerfile               # Agent core container image
├── docker-compose.yml       # Local dev (Ollama GPU/CPU)
└── requirements.txt         # Python dependencies

Quick Start (Local Dev)

# 1. Start Ollama
docker compose --profile gpu up -d

# 2. Pull model
docker exec bms-ollama ollama pull qwen2.5:7b

# 3. Run BMS demo
python -m src --demo

Documentation

• PoC Specification — objectives, requirements, architecture
• Implementation Plan — detailed phased plan with acceptance criteria

Kubernetes Cluster

Namespace	Purpose
shared-services	Ollama (GPU, LLM inference)
bms-ops	BMS application services
db	PostgreSQL
monitoring	Prometheus, Grafana, Tempo, Alertmanager
argocd	GitOps (ArgoCD)
gitea	Git repository (Gitea)
nexus	Container registry (Nexus)