Tensile Strength Predictor

Using XGBoost Model to predict the tensile strength of rock.

Comment

Inspiration

We noticed that civil /mining engineers still rely on slow, expensive lab tests to determine the tensile strength of new rocks. A single test can cost hundreds of dollars and delay product timelines by days. We asked ourselves: “What if a web app could give a reliable estimate in seconds using only easily-measured features like density, porosity, CAI and tensile values?” The hackathon’s “real ML, no wrappers” rule was the perfect push to train our own model and show that small-data, open-source pipelines can still deliver industrial value.

What it does

Tensile Strength Predictor is a full-stack web application that:

  1. Accepts four physical measurements (density, porosity, CAI, tensile) via a clean UI.
  2. Runs a custom-trained XGBoost model (≈ 2 kB) served by FastAPI.
  3. Returns an instant tensile-strength estimate and visual confidence range.
  4. Runs entirely on free cloud tiers (Render + Vercel) with zero LLM APIs.

How we built it

Layer Stack & Flow
Data manually created and tested dataset from rock samples (100+ samples) → cleaned & engineered 6 derived features in a Jupyter notebook.
Model XGBoost regressor tuned with Optuna → exported via joblib → 2 kB model.pkl.
Backend FastAPI endpoint (/predict) reproduces exact feature engineering at inference time → Docker → GHCR → Render free tier.
Frontend Vanilla HTML/CSS/JS single page → fetches backend via HTTPS → deployed on Vercel.
CI/CD GitHub Actions auto-builds & pushes Docker image on every main push.

Everything is MIT-licensed and reproducible from the public repo.

Challenges we ran into

  • CORS: Browser pre-flight OPTIONS requests were blocked locally; solved with FastAPI’s CORSMiddleware.
  • Docker push failures on Windows PowerShell due to Execution-Policy; fixed via Set-ExecutionPolicy RemoteSigned.
  • Render port detection: had to explicitly expose 8000 and set PORT=8000 env var.
  • Model drift risk: ensured training & inference use identical feature engineering code to avoid silent skew.

Accomplishments that we're proud of

  • Sample Collected: Collected 100+ samples and tested physical and mechanical properties of sandstone rock samples in lab.
  • End-to-end pipeline from raw CSV → trained model → live HTTPS API → responsive UI in < 48 h.
  • Sub-100 ms prediction latency on free Render dyno.
  • Full open source: model weights, training code, Dockerfiles and infra-as-code (GitHub Actions) all public.
  • Zero external AI APIs: every line of ML is ours—no black-box calls.

What we learned

  • FastAPI + Docker + Render is a breeze for lightweight Python services.
  • Vercel’s Git-linked deploys let us iterate in minutes, not hours.
  • Feature-engineering parity between training and inference is non-negotiable—one misplaced log transform silently breaks predictions.
  • Free tiers are surprisingly generous for hackathon-scale traffic.

What's next for Tensile Strength Predictor

  • Expand dataset to 10 k+ samples across more alloy families.
  • ONNX export for ultra-low-latency inference (< 5 ms) at plant edge devices.
  • Upload CSV batch mode so engineers can predict hundreds of rows at once.
  • Confidence intervals & SHAP explanations to give engineers trust and insight.
  • Custom domain + SSL + rate limiting for production use in SMEs.

Built With

Share this project:

Updates