Series 7 — Self-Hosted AI Stack on Raspberry Pi • Part 1 of 6
A Raspberry Pi 5 running Kokoro TTS, Whisper STT, Ollama LLM, ChromaDB, and a custom audio converter handles the entire the WhatsApp AI agent stack without a single cloud API call. This article covers the service architecture, port planning, resource management, and startup order.
Services and Ports
┌─────────────────────────────────────────────────────────────────────────────┐
│ RASPBERRY PI 5 (16GB RAM) │
│ │
│ Service Port Language Process manager │
│ ───────────────────────────────────────────────────────── │
│ Kokoro TTS 9010 Python nohup + disown │
│ Whisper STT 9011 Python nohup + disown │
│ Audio Converter 9012 Python nohup + disown │
│ Ollama LLM 11434 Go systemd (built-in) │
│ ChromaDB 8000 Python nohup + disown │
│ Apache (PHP app) 443 PHP/Apache systemd │
│ Nginx (reverse proxy) 80 Nginx systemd │
└─────────────────────────────────────────────────────────────────────────────┘Resource Constraints on Pi 5 (16GB)
| Service | RAM (steady) | RAM (peak) | CPU (idle) |
|---|---|---|---|
| Kokoro TTS (af_heart model) | ~800MB | ~1.2GB | ~2% |
| Whisper STT (base model) | ~300MB | ~600MB | ~1% |
| Ollama (llama3.1:8b) | ~5.5GB | ~6.5GB | ~3% |
| ChromaDB | ~200MB | ~500MB | ~1% |
| Audio Converter | ~50MB | ~200MB | ~0% |
| Apache + PHP | ~400MB | ~800MB | ~5% |
Total steady-state: ~7.25GB. Peak (all services under load simultaneously): ~9.8GB. Leaves 6GB headroom on a 16GB Pi.
Service Startup Order
Services with inter-dependencies must start in order:
- ChromaDB — no dependencies
- Ollama — no dependencies (managed by systemd)
- Kokoro TTS — no dependencies
- Whisper STT — no dependencies
- Audio Converter — depends on FFmpeg being installed
- Apache/PHP — depends on all above (waits for health checks)
Create a startup script that checks each service's health endpoint before starting the next. A service that starts before its dependencies produces cryptic errors.
What to Watch For
- SD card failure — Raspberry Pi running AI services writes frequently to disk (model caches, temp files). Use an SSD via USB 3.0 for the OS and services, not the SD card.
- Temperature throttling — Pi 5 throttles at 80°C. Under sustained LLM load, CPU temperature can reach 70°C without active cooling. Add a heatsink and fan.
- Port conflict detection — Before starting any service, check if the port is already bound:
ss -tlnp | grep :9010. A zombie process holding the port will cause a silent startup failure.