WolfDavid/ai-internet-diagnostic-model

AI Internet Diagnostic

Tells you the specific reason your Wi-Fi just dropped — evidence-grounded, confidence-scored attribution like "your school's 802.1X session expired at 09:14:23 — here are the three telemetry signals that prove it."

Results

Macro F1: 0.974 (synthetic) · pending (real, Reality Anchor dogfood) · ECE 0.28

Architecture

Mermaid source (renders on GitHub)

flowchart LR
  L["📡 Laptop telemetry"] --> S["📋 wifi-diag-schema"]
  S --> CLS["🔢 LightGBM 10-class classifier"]
  S --> ANO["📈 PyOD IForest anomaly detector"]
  CLS --> V["📊 Verdict + EvidenceItems"]
  ANO --> V
  V --> N["💬 Anthropic Haiku 4.5 narrator"]
  N --> UI["🖥️ Gradio Live tab + Agent CLI"]
  style CLS fill:#3498db,stroke:#1b4f72,stroke-width:3px,color:#fff
  style ANO fill:#3498db,stroke:#1b4f72,stroke-width:3px,color:#fff
  style V fill:#2ecc71,stroke:#196f3d,stroke-width:2px,color:#fff

Trained models (blue) sit at the visual gravity center of the pipeline. The LLM narrator (green) is downstream — it explains what the classifier and anomaly detector found, with citations to specific telemetry fields. This is not a GPT wrapper.

Try it live

🔗 Live demo on Hugging Face Spaces

---

AI Internet Diagnostic — Model Repo

LightGBM 10-class disconnect classifier + PyOD anomaly detector + reproducible synthetic-data generator for the AI Internet Diagnostic project.

This is one of four repos in the project topology (D-10 / D-11):

• ai-internet-diagnostic-space — Hugging Face Space (Phase 3)
• ai-internet-diagnostic-model (this repo) — model artefacts + synthetic-data generator (Phase 1–2)
• ai-internet-diagnostic-agent — cross-platform local telemetry agent (Phase 4)
• wifi-diag-schema — Pydantic wire-format schema, published to PyPI (Phase 1)

Quickstart

uv sync --all-extras --dev
make synth     # regenerate data/train.parquet (100k) + data/eval.parquet (20k); <30s
make test      # run unit tests

Reproducibility

make synth regenerates train + eval Parquet byte-identically from fixed master seeds (D-08):

• MASTER_TRAIN_SEED = 20260501 → data/train.parquet (10,000 samples × 10 classes)
• MASTER_EVAL_SEED = 20260502 → data/eval.parquet (2,000 samples × 10 classes)

Per-class PCG64 sub-streams via SeedSequence.spawn() (RESEARCH Pattern 4) guarantee determinism.

Datasheet

See DATASHEET.md for the Gebru-format dataset card. Per CONTEXT.md D-09, the Limitations section leads with the synthetic-vs-real gap; the Reality Anchor placeholder is reserved for Phase 4 dogfood data.

License

Apache-2.0