Ethica — Ethical Code Analysis Powered Entirely by Gemma 4 (Raspberry Pi Setup)

This is a submission for the Gemma 4 Challenge: Build with Gemma 4

What I Built

Ethica — a pure-AI ethical code analysis tool that uses Gemma 4 as the sole ethics auditor. No external static analysis tools, no pre-written rules — Gemma 4 reads code, identifies bias, accessibility issues, security concerns, and ethical risks through pure reasoning.

The tool runs entirely on-device on a Raspberry Pi 5 (8GB RAM) with a quantized Gemma 4 GGUF model. It features hierarchical analysis to handle large codebases within context limits, splitting files into function-level chunks and summarizing where needed.

Demo

The tool is CLI-based. Here's a sample run detecting gender bias in code:

$ echo 'def process_users(users):
    for user in users:
        if user.gender == "M":
            send_mail(user, "Dear Mr.")
        else:
            send_mail(user, "Dear Mrs.")' | python run_cli.py analyze --stdin --dimension bias

╭─────────────────────────────────── ETHICA - BIAS Analysis ───────────────────────────╮
│ Issue: Using binary gender classification ("M" vs "F") to determine salutation...     │
│ Severity: MEDIUM                                                                       │
│ Recommendations:                                                                       │
│   - Use gender-neutral salutation ("Dear User")                                       │
│   - Add fallback for null/unknown gender values                                       │
│   - Audit dataset for systemic biases                                                 │
╰───────────────────────────────────────────────────────────────────────────────────────╯

Code

Repository: https://github.com/ether-btc/gemma-4-challenge

git clone https://github.com/ether-btc/gemma-4-challenge.git
cd gemma-4-challenge
python -m venv venv && source venv/bin/activate && pip install -r requirements.txt

# Test
PYTHONPATH=src:. python run_cli.py test

# Analyze a file
PYTHONPATH=src:. python run_cli.py analyze path/to/code.py --dimension bias

# Analyze via stdin
echo 'code here' | PYTHONPATH=src:. python run_cli.py analyze --stdin --dimension security

How I Used Gemma 4

Model: Gemma 4 E2B (2B parameters, quantized to Q8_0 GGUF, 4.6GB)

Why E2B: The Raspberry Pi 5 has limited RAM (8GB). E2B is the only Gemma 4 variant that fits alongside the operating system and other tooling without OOM. The quantized GGUF runs via llama-cpp-python with the pi-5 branch optimizations.

How it works:

1. User provides code via file or stdin
2. GemmaClient loads the GGUF model and crafts dimension-specific prompts (bias, accessibility, security, ethics)
3. For large files, HierarchicalAnalyzer splits by function using regex, analyzes each chunk, and combines results with confidence scoring
4. Gemma 4 performs all ethical reasoning — no rules, no patterns, just pure LLM analysis
5. Results output as formatted panels (CLI), JSON, or Markdown

Prompt architecture:

• System prompts specialize Gemma 4 as an ethical code reviewer per dimension
• Structured output format requests JSON with issues, severity, recommendations, confidence
• Fallback regex parser handles non-JSON responses gracefully

The tool respects the contest's "pure AI-native" constraint: all analysis is performed by Gemma 4 reasoning, no external tools involved.