Christopher — Local Voice AI Assistant

Fully local. Privacy-first. Real-time. Built from source.

🌐 fusional.dev — Production MCP deployments by the team behind Christopher

Christopher is a fully offline voice AI pipeline running entirely on your hardware. No cloud. No API keys. No data leaving your machine. You speak — it listens, thinks, and talks back.

Built on three compiled-from-source inference engines:

• whisper.cpp — speech recognition
• llama.cpp — LLM inference with CUDA GPU acceleration
• Piper TTS — neural text-to-speech

---

Stack

Component	Engine	Model
Speech Recognition	whisper.cpp	ggml-base.en
Language Model	llama.cpp + CUDA	Model profile: Llama 3.2 3B / Qwen2.5 3B / Mistral 7B
Text to Speech	Piper	en_US-libritts-high

Hardware: i5-7300HQ + GTX 1050 Ti 4GB + WSL2 Ubuntu

---

How It Works

Microphone → whisper.cpp → transcript → llama.cpp → reply → Piper TTS → speakers

One bash script. Three binaries. Zero cloud dependencies.

---

Quick Start (Pilot Installer)

Single command — checks all prerequisites first, then builds and downloads everything:

bash pilot_install.sh

After install, run the health check and launch Christopher:

bash preflight_voice.sh       # verify the pipeline is ready
python3 christopher.py --voice  # start voice mode

See docs/pilot-setup-guide.md for supported host configurations, failure recovery, and full walkthrough. See docs/offline-runbook.md for offline startup, fallback behaviors, and air-gapped deployment.

---

Manual Requirements

# whisper.cpp
git clone https://github.com/ggerganov/whisper.cpp
cd whisper.cpp && mkdir build && cd build
cmake .. && make -j4
bash models/download-ggml-model.sh base.en

# llama.cpp with CUDA
git clone https://github.com/ggerganov/llama.cpp
cd llama.cpp && mkdir build && cd build
cmake .. -DGGML_CUDA=ON && make -j4
# Default install downloads Llama 3.2 3B Q4_K_M into models/

# Piper TTS
pip install piper-tts

# Christopher orchestrator deps
pip install -r christopher_requirements.txt

# System
sudo apt install alsa-utils ffmpeg

---

Usage

chmod +x voice_ai.sh
chmod +x preflight_voice.sh
./preflight_voice.sh
./voice_ai.sh

Speak after the prompt. Christopher listens for 5 seconds, generates a response, speaks it back.

WSL2 Microphone Troubleshooting

If voice mode starts but never detects speech:

1. Start PulseAudio on Windows (C:\PulseAudio\start-pulseaudio.cmd).
2. In WSL, set PULSE_SERVER if needed:

export PULSE_SERVER=tcp:$(awk '/^nameserver / {print $2; exit}' /etc/resolv.conf):4713

3. Re-run ./voice_ai.sh.

The script now tries PulseAudio capture first (parec) and falls back to ALSA (arecord) automatically.

---

GPU Tuning

Christopher now uses a model-profile system so you can switch models without editing multiple files.

VRAM	Recommended flags
4GB (GTX 1050 Ti)	llama32-3b or qwen25-3b with profile defaults
4GB (GTX 1050 Ti, slower)	mistral-7b with -ngl 28 -c 512
8GB	-ngl 40 -c 2048
16GB+	-ngl 99 -c 4096

Model Profiles

Christopher can now run three named local model profiles:

• llama32-3b — default; best fit for 4GB VRAM
• qwen25-3b — best candidate if you want a smarter small model on the same hardware
• mistral-7b — usable, but slower on 4GB VRAM

Set the default in .env:

MODEL_PROFILE=llama32-3b

Or switch per run:

python3 christopher.py --chat --model-profile qwen25-3b
python3 christopher.py --voice --model-profile llama32-3b
python3 christopher.py --chat --model-profile mistral-7b --ngl 28 --ctx 512

Benchmarking Models

Use the built-in benchmark mode to compare latency before adopting a model:

python3 christopher.py --benchmark --model-profile llama32-3b
python3 christopher.py --benchmark --model-profile qwen25-3b
python3 christopher.py --benchmark --model-profile mistral-7b --ngl 28 --ctx 512

This prints per-run latency and the average so you can compare on your actual hardware.

---

Roadmap

• Wake word detection
• MCP tool integration (voice control over FusionAL — database, Slack, GitHub)
• Centralized model profiles + benchmark mode
• Test Qwen2.5 3B against Llama 3.2 3B on GTX 1050 Ti
• RAG module for local document search
• WebSocket streaming
• GUI dashboard

---

Origin

This was my first ever systems-level build.

I built Christopher in a hospital over 4-5 days — compiling whisper.cpp and llama.cpp from source, getting CUDA running on a GTX 1050 Ti, wiring the full pipeline in bash. No prior C++ experience. No prior CUDA experience. Just time, necessity, and refusal to stop.

The name Christopher belongs to my brother and his son. His son saved me. This project is dedicated to them.

---

Author

Jonathan Melton — Self-taught developer. Privacy-first AI engineering.

Projects:

• FusionAL — Self-hosted MCP governance gateway
• mcp-consulting-kit — Production MCP servers
• Christopher-AI — This project

Consulting: fusional.dev — Done-for-you MCP governance deployments ($3.5k-9k)
Contact: jonathanmelton.fusional@gmail.com • Book Call