% ============================================================================
% TikZ: Robust tabular neural network architecture
% ============================================================================
\begin{tikzpicture}[
  font=\small,
  >=Latex,
  node distance=6mm and 6mm,
  box/.style={rounded corners=2pt, draw=black!55, align=center,
              minimum height=8mm, inner sep=3pt, text width=24mm},
  cat/.style={box, fill=NavyBlue!8},
  num/.style={box, fill=OliveGreen!10, text width=30mm},
  emb/.style={box, fill=BurntOrange!14, text width=22mm},
  op/.style={box, fill=black!5},
  spine/.style={box, fill=black!6, text width=52mm},
  res/.style={box, fill=Plum!10, text width=52mm, minimum height=11mm},
  outb/.style={box, fill=NavyBlue!16, text width=34mm, font=\bfseries},
  detail/.style={rounded corners=2pt, draw=black!40, dashed, fill=Plum!5,
                 align=left, text width=40mm, inner sep=4pt, font=\footnotesize},
  ar/.style={->, thick, black!65}
]

% ---- inputs ----
\node[cat] (city) {City index\\{\footnotesize (195 cats)}};
\node[cat, right=of city] (type) {Type index};
\node[cat, right=of type] (src) {Source index};
\node[num, right=10mm of src] (numin) {Numeric features\\{\footnotesize $\approx$45: area, lot, beds,\\baths, age, amenities,\\rooms, GPT specs}};

% ---- embeddings / bn ----
\node[emb, below=of city] (ecity) {Embedding\\{\footnotesize dim 29}};
\node[emb, below=of type] (etype) {Embedding\\{\footnotesize dim 4}};
\node[emb, below=of src] (esrc) {Embedding\\{\footnotesize dim 2}};
\node[op, below=of numin, text width=30mm] (bn) {BatchNorm\\{\footnotesize + missingness flags}};

\draw[ar] (city)--(ecity); \draw[ar] (type)--(etype);
\draw[ar] (src)--(esrc);  \draw[ar] (numin)--(bn);

% ---- concat ----
\node[spine, below=12mm of etype.south, xshift=8mm] (concat)
  {Concatenate \;$\rightarrow$\; $\mathbb{R}^{\,\text{emb}+\text{num}}$};
\draw[ar] (ecity) |- (concat.west);
\draw[ar] (etype) -- (concat);
\draw[ar] (esrc)  -- (concat);
\draw[ar] (bn)    -| (concat.east);

% ---- projection ----
\node[spine, below=of concat] (proj) {Linear $\rightarrow$ 256\;\; (trunk)};
\draw[ar] (concat)--(proj);

% ---- residual blocks ----
\node[res, below=of proj] (resb)
  {Residual block \;$\times\,3$\\{\footnotesize BN $\to$ GELU $\to$ Lin$_{512}$ $\to$ Dropout $\to$ Lin$_{256}$ \;$\oplus$\; skip}};
\draw[ar] (proj)--(resb);

% ---- head ----
\node[spine, below=of resb] (head)
  {Head:\; BatchNorm $\to$ GELU $\to$ Linear$_{1}$};
\draw[ar] (resb)--(head);

% ---- output ----
\node[outb, below=of head] (out) {$\widehat{\ln(\text{Price})}$};
\draw[ar] (head)--(out);

% ---- residual detail callout ----
\node[detail, right=10mm of resb] (dtl)
  {\textbf{Res block:} $x \mapsto x + f(x)$\\
   $f=$ Drop$\circ$Lin$_{256}\!\circ$Drop$\circ$GELU\\ $\circ$Lin$_{512}\!\circ$BN};
\draw[->, dashed, black!45] (resb.east)--(dtl.west);

% ---- training / ensemble annotation ----
\node[detail, right=10mm of head, text width=42mm, fill=OliveGreen!6, draw=black!40] (train)
  {\textbf{Training (robust):}\\
   Huber loss $\cdot$ AdamW $\cdot$ cosine LR\\
   grad-clip 5 $\cdot$ dropout $\cdot$ early stop};

% ---- ensemble + blend ----
\node[op, below=8mm of out, text width=64mm, fill=NavyBlue!7] (ens)
  {Ensemble of 3 seeds (average) \;$\Rightarrow$\; blend 50/50 with gradient boosting};
\draw[ar] (out)--(ens);

\end{tikzpicture}