Setup for BOxCrete

Dockerfile

@@ -0,0 +1,17 @@
+FROM python:3.11-slim
+
+WORKDIR /app
+
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    build-essential git \
+    && rm -rf /var/lib/apt/lists/*
+
+COPY requirements.txt .
+RUN pip install --no-cache-dir -r requirements.txt
+
+COPY main.py .
+
+# HF Spaces requires port 7860
+EXPOSE 7860
+
+CMD ["python", "main.py"]

README.md

@@ -1,10 +1,34 @@
 ---
-title: Mix
-emoji: 🐨
-colorFrom: indigo
-colorTo: pink
+title: MetaMix API
+emoji: 🏗️
+colorFrom: gray
+colorTo: blue
 sdk: docker
+app_port: 7860
 pinned: false
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# MetaMix API
+
+FastAPI backend for the [MetaMix](https://pavements.design/metamix) tool on pavements.design.
+
+Wraps [BOxCrete](https://github.com/facebookresearch/SustainableConcrete) — a Bayesian
+optimisation framework for sustainable concrete mix design by Meta AI Research.
+
+## Endpoints
+
+- `GET /health` — liveness check
+- `POST /predict` — returns GWP and strength curve for a given mix
+
+## Citation
+
+```
+@misc{ament2023sustainable,
+      title={Sustainable Concrete via Bayesian Optimization},
+      author={Sebastian Ament and Andrew Witte and Nishant Garg and Julius Kusuma},
+      year={2023},
+      eprint={2310.18288},
+      archivePrefix={arXiv},
+      primaryClass={cs.LG}
+}
+```

main.py

@@ -0,0 +1,186 @@
+"""
+MetaMix API — FastAPI wrapper around BOxCrete
+=============================================
+Designed to run as a Hugging Face Space (Docker SDK).
+
+Deploy steps:
+  1. Create a new Space at https://huggingface.co/spaces
+     - Owner: your HF username
+     - Space name: metamix-api  (or whatever you like)
+     - SDK: Docker
+  2. Push this file + requirements.txt + README.md to the Space repo
+  3. HF builds the Docker image automatically
+  4. Copy the Space URL into Vercel env: METAMIX_API_URL=https://<user>-metamix-api.hf.space/predict
+
+Cold start note:
+  The model fits from the bundled BOxCrete dataset on first request (~30 s).
+  Subsequent requests in the same session are fast (< 2 s).
+  HF free Spaces sleep after ~15 min of inactivity — the next request will
+  trigger another cold start. The page.tsx loading state handles this gracefully.
+"""
+
+from __future__ import annotations
+
+import logging
+from contextlib import asynccontextmanager
+from typing import Optional
+
+import torch
+from fastapi import FastAPI, HTTPException
+from fastapi.middleware.cors import CORSMiddleware
+from pydantic import BaseModel, Field
+
+from boxcrete.models import SustainableConcreteModel
+from boxcrete.utils import load_concrete_strength, get_bounds
+
+logger = logging.getLogger(__name__)
+
+# ── Global model state ────────────────────────────────────────────────────────
+
+_model: Optional[SustainableConcreteModel] = None
+_data = None
+
+CURVE_DAYS = list(range(1, 91))
+STRENGTH_DAYS = [1, 3, 7, 14, 28, 56, 90]
+
+
+def _fit_model() -> SustainableConcreteModel:
+    global _data
+    logger.info("Loading BOxCrete dataset…")
+    _data = load_concrete_strength()
+    _data.bounds = get_bounds(_data.X_columns)
+
+    model = SustainableConcreteModel(strength_days=STRENGTH_DAYS)
+    logger.info("Fitting GWP model…")
+    model.fit_gwp_model(_data)
+    logger.info("Fitting strength model…")
+    model.fit_strength_model(_data)
+    logger.info("Model ready.")
+    return model
+
+
+@asynccontextmanager
+async def lifespan(app: FastAPI):
+    global _model
+    _model = _fit_model()
+    yield
+
+
+# ── App ───────────────────────────────────────────────────────────────────────
+
+app = FastAPI(title="MetaMix API", lifespan=lifespan)
+
+app.add_middleware(
+    CORSMiddleware,
+    allow_origins=[
+        "https://pavements.design",
+        "https://www.pavements.design",
+        "http://localhost:3000",
+    ],
+    allow_methods=["POST", "GET"],
+    allow_headers=["Content-Type"],
+)
+
+# ── Schemas ───────────────────────────────────────────────────────────────────
+
+class MixRequest(BaseModel):
+    cement: float = Field(..., ge=0, le=600)
+    fly_ash: float = Field(0.0, ge=0, le=400)
+    slag: float = Field(0.0, ge=0, le=400)
+    water: float = Field(..., ge=80, le=350)
+    fine_aggregate: float = Field(0.0, ge=0, le=1200)
+    coarse_aggregate: float = Field(0.0, ge=0, le=1200)
+    superplasticizer: float = Field(0.0, ge=0, le=30)
+
+
+class StrengthPoint(BaseModel):
+    day: int
+    mean: float
+    lower: float
+    upper: float
+
+
+class PredictionResponse(BaseModel):
+    gwp: float
+    gwp_lower: float
+    gwp_upper: float
+    strength_28d: float
+    strength_28d_lower: float
+    strength_28d_upper: float
+    strength_curve: list[StrengthPoint]
+
+
+# ── Helpers ───────────────────────────────────────────────────────────────────
+
+def _build_comp_tensor(req: MixRequest, comp_cols: list[str]) -> torch.Tensor:
+    """Map request fields onto the column order BOxCrete expects."""
+    mapping = {
+        "Cement (kg/m3)": req.cement,
+        "Fly Ash (kg/m3)": req.fly_ash,
+        "Slag (kg/m3)": req.slag,
+        "Water (kg/m3)": req.water,
+        "Fine Aggregate (kg/m3)": req.fine_aggregate,
+        "Coarse Aggregate (kg/m3)": req.coarse_aggregate,
+        "Superplasticizer (kg/m3)": req.superplasticizer,
+    }
+    values = [mapping.get(col, 0.0) for col in comp_cols]
+    return torch.tensor(values, dtype=torch.float64).unsqueeze(0)  # 1 × d
+
+
+def _posterior_stats(model, X: torch.Tensor) -> tuple[float, float, float]:
+    with torch.no_grad():
+        posterior = model.posterior(X)
+        mean = posterior.mean.squeeze().item()
+        std = posterior.variance.squeeze().item() ** 0.5
+    return mean, mean - 1.96 * std, mean + 1.96 * std
+
+
+# ── Endpoints ─────────────────────────────────────────────────────────────────
+
+@app.get("/health")
+async def health():
+    return {"status": "ok", "model_ready": _model is not None}
+
+
+@app.post("/predict", response_model=PredictionResponse)
+async def predict(req: MixRequest):
+    if _model is None or _data is None:
+        raise HTTPException(status_code=503, detail="Model not yet initialised")
+
+    # Composition columns (everything except Time)
+    comp_cols = _data.X_columns[:-1]
+    comp = _build_comp_tensor(req, comp_cols)  # 1 × (d-1)
+
+    # GWP
+    gwp_mean, gwp_lo, gwp_hi = _posterior_stats(_model.gwp_model, comp)
+
+    # Strength curve (day 1 → 90)
+    curve: list[StrengthPoint] = []
+    str28_mean = str28_lo = str28_hi = 0.0
+
+    for day in CURVE_DAYS:
+        t = torch.tensor([[float(day)]], dtype=torch.float64)
+        X_t = torch.cat([comp, t], dim=-1)
+        mean, lo, hi = _posterior_stats(_model.strength_model, X_t)
+        mean = max(mean, 0.0)
+        lo = max(lo, 0.0)
+        curve.append(StrengthPoint(day=day, mean=round(mean, 2), lower=round(lo, 2), upper=round(hi, 2)))
+        if day == 28:
+            str28_mean, str28_lo, str28_hi = mean, lo, hi
+
+    return PredictionResponse(
+        gwp=round(gwp_mean, 2),
+        gwp_lower=round(gwp_lo, 2),
+        gwp_upper=round(gwp_hi, 2),
+        strength_28d=round(str28_mean, 2),
+        strength_28d_lower=round(str28_lo, 2),
+        strength_28d_upper=round(str28_hi, 2),
+        strength_curve=curve,
+    )
+
+
+# ── Entry point (HF Spaces requires port 7860) ────────────────────────────────
+
+if __name__ == "__main__":
+    import uvicorn
+    uvicorn.run(app, host="0.0.0.0", port=7860)

requirements.txt

@@ -0,0 +1,5 @@
+fastapi[standard]>=0.115.0
+uvicorn>=0.30.0
+# CPU-only PyTorch — smaller image, sufficient for GP inference
+torch>=2.2.0 --index-url https://download.pytorch.org/whl/cpu
+git+https://github.com/facebookresearch/SustainableConcrete.git