Mobile · Mind Matrix (internship) · 2026

Road Health Tracker

Mobile app that classifies road condition from phone sensor data and maps damaged segments.

Android app (internship) using computer vision and generative AI to flag road condition, with on-device storage. Rebuilt as a public demo: FastAPI + scikit-learn classifier on accelerometer features, map view, history and reports.

My role: Android intern (original app); sole engineer (public rebuild)
Links: GitHub ↗
Stack: AndroidAndroid StudioPythonFastAPIscikit-learn; Android · Kotlin/Java · Room · GenAI · FastAPI · scikit-learn · SQLite
Microsoft-ecosystem fit: ▸ Model portable to ONNX Runtime (planned)

5/5

API tests passing

0.956

Macro-F1 — synthetic sample, not real roads

[REPLACE]

Real-world accuracy / test drives

Source · Generated by the real API: reports logged via POST /reports (synthetic sensor features), its live /docs, and the classifier's hold-out confusion matrix. [REPLACE] add emulator screenshots of the app

01 — Problem

Road damage reporting was manual and slow, although smartphones already capture the sensor data needed to detect it.

02 — Approach

01
2-second accelerometer/gyro windows → features (vertical RMS, peak, variance, speed).
02
Random-forest classifier: good / fair / poor, with confidence.
03
Reports stored locally first, synced later; map + history + summary report screens.

03 — Architecture

System architecture

S01
Phone sensors
S02
Feature window
S03
Classifier
S04
SQLite / Room
S05
Map · History · Report

04 — Engineering notes

From a vibration window to a label

The public rebuild classifies a 2-second sensor window. The features are cheap enough to compute on the phone, so the app can label roads offline and sync later.

python

# services/road-health-api/app/model.py
FEATURES = ["z_rms", "z_peak", "z_std", "gyro_var", "speed_kmh"]
LABELS = ["good", "fair", "poor"]

def predict(features: dict) -> tuple[str, float]:
    x = np.array([[features[f] for f in FEATURES]])
    proba = load().predict_proba(x)[0]
    i = int(proba.argmax())
    return LABELS[i], float(proba[i])

Honest numbers

The classifier scores 0.956 macro-F1 on a held-out split of a synthetic dataset, generated with physically plausible vibration ranges. That proves the pipeline, not real-world accuracy; real labelled drives replace make_dataset().