Offline Windows event log triage. It flags attacker activity and explains each finding in plain English, using small models that run on your own machine.
Submitted to the Ryze AI hackathon (Small Models track). Every model is 500M parameters or fewer and runs locally.
My background is digital forensics and pulling evidence off devices. Most log triage tools assume you already know what you're looking at. I wanted something simple that runs locally, takes a couple of log files, and tells you what matters and why.
Drop in .evtx files from a Windows machine and get back the events a responder looks at first:
| Event ID | Log | What it means |
|---|---|---|
| 4625 | Security | Failed logon. Bursts from one source suggest brute force. One source hitting many accounts suggests password spraying. |
| 7045 | System | A new service was installed. Attackers use services for persistence and remote execution (PsExec, etc). |
| 1102 | Security | The Security log was cleared. Almost never legitimate on a production machine. |
Event logs are full of usernames, hostnames and IP addresses, and during an incident the machine may be offline or untrusted. Nothing here sends data anywhere.
git clone https://github.com/sqlinjected/evtx
cd evtx
python -m venv .venv
.venv\Scripts\activate
pip install -e ".[ml,dev]"
evtx-triage --demo # no logs needed
evtx-triage Security.evtx System.evtx # real logs
evtx-triage --eval # model vs rules
On macOS/Linux activate with source .venv/bin/activate. Add --json for JSON output, or --no-model for keyword rules only (you can also drop ml from the install).
The first run with the model downloads about 90 MB of weights to ~/.cache/evtx-triage. After that it works fully offline.
Don't run attack simulations on your real machine. Use a throwaway Windows VM and run scripts/generate_test_events.ps1 from an elevated PowerShell. It produces all three event types and exports the logs to samples/.
I use all-MiniLM-L6-v2 (22M parameters, Apache-2.0), with the ONNX export from Qdrant, running locally through ONNX Runtime via fastembed. No external API calls.
The keyword rules only catch a service install when its path contains a known bad string like \Users\Public\ or powershell. The model is a second opinion on every install.
- Each install path gets a few plain-English notes: where the file lives (Program Files, the system folder, a user profile, a temp folder, a network share), how it launches (command shell, PowerShell, rundll32), and whether the service name looks randomly generated.
- MiniLM embeds the service name plus those notes. I leave the raw path out. When I tested with raw paths, MiniLM latched onto surface details like quotes, length and version numbers, and accuracy dropped from 30/35 to 21/35.
- Each embedding is compared against the average of two short reference lists in
anomaly.py: ordinary software installs, and known attacker tradecraft.