AILog — AI Log Triage for AOSP & Android Automotive

Stop drowning in 50,000 log lines. Let AILog find what matters.

AILog reads your AOSP build errors, adb logcat, and full bugreports — and tells you what actually broke and how to fix it. It's built for AOSP and Android Automotive (AAOS) platform developers debugging VHAL, CarService, HALs, and native crashes in a terminal.

• 🔒 Local-first — runs on Ollama by default, so your logs never leave your machine
• 🚗 Automotive-aware — ships a knowledge pack of VHAL, SELinux, tombstone & CarService facts
• 🪶 Zero dependencies — pure Python standard library, installs anywhere
• ⚡ Works offline — instant rule-based triage even with no AI model at all

pip install ailog-cli

Note: the PyPI package is ailog-cli, but the command you run is ailog. (The name ailog on PyPI belongs to an unrelated project — don't install that one.)

See it in action

Turn a 100 MB bugreport into a ranked list of real problems — instantly, with no model needed:

Try it yourself on the bundled sample — no device required:

git clone https://github.com/zoddiacc/ailog-cli.git
ailog bugreport ailog-cli/docs/demo/bugreport-car-demo.zip --no-ai

Add a model and it explains each crash in depth. Point it at a live device (ailog cat --explain) or your AOSP build (ailog build) for the same treatment.

Get started in 60 seconds

1. Install — the package is ailog-cli, the command is ailog:

pip install ailog-cli

2. Try it with zero setup — the knowledge pack works with no AI at all:

ailog bugreport your-bugreport.zip --no-ai

3. Turn on AI — pick one:

# Install Ollama once: https://ollama.com
ollama pull qwen2.5-coder:3b     # ~2 GB, one-time download

ailog config --provider openai --api-key sk-...
# or
ailog config --provider anthropic --api-key sk-ant-...
# or any OpenAI-compatible endpoint:
ailog config --provider openai --base-url https://api.groq.com/openai/v1 --api-key ...

4. Use it

ailog analyze build.log          # analyze a saved log
ailog cat --explain              # live logcat, explained inline
ailog build                      # wrap an AOSP build
ailog bugreport report.zip       # triage a bugreport

Run ailog --help or ailog <command> --help for everything.

Why AILog is different

Most people run a small local model (the default is qwen2.5-coder:3b) that knows almost nothing about VHAL, CarService, SELinux, or tombstones. A generic "pipe the log to an LLM" tool gives weak or wrong answers for automotive internals as a result.

AILog keeps the domain intelligence in curated data it ships, not the model's weights. A built-in knowledge pack maps log signatures to verified facts — so even a tiny local model gives genuinely good answers, and the common cases are explained instantly with no AI at all.

What you can do

# Analyze
ailog analyze logcat.txt --focus CarService      # focus the AI on a component
ailog analyze build.log --output report.md       # save a markdown report

# Bugreport triage
ailog bugreport report.zip --no-ai               # instant, offline, no model
ailog bugreport report.zip --focus com.oem.app   # only issues touching a package

# Live logcat
ailog cat --explain                              # explain each error inline
ailog cat --focus VHAL --noise-level high        # focus + aggressive filtering
ailog cat -p com.example.app --explain           # filter to one app

# AOSP build (run from a lunch'd shell)
ailog build
ailog build -- -j16 framework

# Machine-readable output for CI
ailog --json bugreport report.zip --no-ai | jq '.issue_counts'

Configuration

AILog works out of the box with Ollama — you only need config to switch providers or tune behavior.

ailog config --show                       # see current settings
ailog config --provider anthropic         # switch provider
ailog config --model qwen2.5-coder:7b     # pick a model
ailog config --list-models                # list installed Ollama models
ailog config --set noise_level=high       # set any option (below)
ailog config --reset                      # back to defaults

Config lives at ~/.config/ailog/config.json (created with 0600 permissions). API keys can also come from OPENAI_API_KEY / ANTHROPIC_API_KEY env vars, which take precedence.

Privacy

With local Ollama (the default) nothing leaves your machine — ideal for OEM and Tier-1 environments. When you use a cloud provider, secrets (API keys, tokens, passwords, JWTs, …) are redacted from log content and source files by default before anything is sent. Pass --no-redact only if you understand the implications.

Requirements

• Python 3.9+ (standard library only — no pip dependencies)
• adb for ailog cat — from Android SDK Platform Tools
• Ollama for local AI (ollama.com), or an API key for cloud AI

analyze and bugreport need only a file — no device or adb.

git clone https://github.com/zoddiacc/ailog-cli.git && cd ailog-cli
bash install.sh                 # installs to ~/.local/bin
# or run directly without installing:
python3 run.py --help

echo 'export PATH="$HOME/.local/bin:$PATH"' >> ~/.zshrc && source ~/.zshrc

How it works

Input (log file, adb logcat, bugreport, or build output)
        │
Stage 1 · Rule-based noise filter    → drops ~70% of lines (instant, free)
        │
Stage 2 · Knowledge-pack lookup      → instant hint (no AI) + facts for the AI
        │
Stage 3 · AI analysis (if needed)    → explains the rest, grounded in Stage 2
        │
Terminal output (color-coded lines, boxed analysis, stats)

The knowledge pack

The heart of AILog. It maps log signatures to verified AOSP/Automotive facts (what a VHAL property means and which Car.PERMISSION_* it needs, how to read an avc: denied line, what a SIGABRT tombstone implies), used two ways:

1. Instant hints, no AI — a matching line gets an always-correct one-liner immediately. This is why ailog bugreport --no-ai is genuinely useful offline.
2. Grounded AI answers — the matching facts are injected into the prompt as authoritative context, so a small local model summarizes known-good knowledge instead of guessing.

It currently covers 50+ error signatures and 120+ VHAL properties (powertrain, energy/EV, HVAC, body, lights, wipers, ADAS, power/user HAL, watchdog, cluster, diagnostics). It's pure data, so it grows without code changes — and tools/gen_vhal_knowledge.py can generate VHAL entries straight from a VehicleProperty.aidl in an AOSP tree.

Contributing

Contributions are welcome — especially new knowledge-pack entries (VHAL properties, SELinux/CarService/build signatures), which are pure data and a great first PR. See CONTRIBUTING.md for setup, and SECURITY.md to report vulnerabilities privately.

python3 -m unittest discover -s tests -v     # run the test suite
ruff check src/ tests/ tools/                # lint

Tests run in CI across Python 3.9–3.14 with ruff linting.

Further reading

• ARCHITECTURE.md — design and internals
• TESTING.md — testing against a real Android device
• CONTRIBUTING.md · CHANGELOG.md · UNINSTALL.md

License

MIT