Create app.py

app.py

@@ -0,0 +1,71 @@
+import os
+import pandas as pd
+import gradio as gr
+from huggingface_hub import hf_hub_download
+from llama_cpp import Llama                         # GGUF inference on CPU
+
+# ---------- model loading (done once at startup) ----------
+MODEL_REPO  = "TheBloke/phi-2-GGUF"                 # fully open 2.7 B model
+MODEL_FILE  = "phi-2.Q4_K_M.gguf"                  # 4‑bit, 3.5 GB RAM
+CTX_SIZE    = 2048                                  # ample for prompt+answer
+
+model_path = hf_hub_download(repo_id=MODEL_REPO, filename=MODEL_FILE)
+llm = Llama(model_path=model_path,
+            n_ctx=CTX_SIZE,
+            n_threads=os.cpu_count() or 2)          # use all CPUs
+
+# ---------- analysis + generation ----------
+def analyze_ads(file):
+    df = pd.read_csv(file.name)
+
+    req = {"headline","description","impressions","CTR","form_opens","spend"}
+    if not req.issubset(df.columns):
+        return f"Missing columns: {', '.join(req - set(df.columns))}"
+
+    # numeric conversions
+    for col in ["impressions","CTR","form_opens","spend"]:
+        df[col] = pd.to_numeric(df[col], errors="coerce")
+    df = df.dropna()
+
+    df["engagement_rate"]     = df["form_opens"] / df["impressions"]
+    df["CPC"]                 = df["spend"] / (df["CTR"] * df["impressions"]).replace(0, pd.NA)
+    df["cost_per_form_open"]  = df["spend"] / df["form_opens"].replace(0, pd.NA)
+
+    top  = df.sort_values("CTR", ascending=False).head(3)
+    worst = df.sort_values("CTR").head(3)
+
+    def rows_to_text(sub):
+        out = ""
+        for _, r in sub.iterrows():
+            out += (f"Headline: {r.headline}\n"
+                    f"Description: {r.description}\n"
+                    f"Imp: {int(r.impressions)}, CTR: {r.CTR:.3f}, "
+                    f"Form Opens: {int(r.form_opens)}, ER: {r.engagement_rate:.3f}\n"
+                    f"Spend: ${r.spend:.2f}, CPC: ${r.CPC:.2f}, CPF: ${r.cost_per_form_open:.2f}\n\n")
+        return out
+
+    prompt = (
+        "You are a senior digital marketer.\n"
+        "Analyse the high‑ and low‑performing ads below and deliver:\n"
+        "1. Key patterns of winners.\n"
+        "2. Weak points of losers.\n"
+        "3. Three actionable creative improvements.\n\n"
+        f"--- HIGH CTR ADS ---\n{rows_to_text(top)}"
+        f"--- LOW CTR ADS ---\n{rows_to_text(worst)}"
+    )
+
+    # generate (stream=False -> returns dict)
+    answer = llm(prompt, max_tokens=320, temperature=0.7, top_p=0.9)["choices"][0]["text"]
+    return answer.strip()
+
+# ---------- Gradio UI ----------
+demo = gr.Interface(
+    fn=analyze_ads,
+    inputs=gr.File(label="CSV with: headline, description, impressions, CTR, form_opens, spend"),
+    outputs=gr.Textbox(label="AI‑generated analysis & recommendations"),
+    title="Ad Performance Analyzer (Phi‑2 4‑bit, CPU‑only)",
+    description="Upload your ad data and get actionable insights without paid APIs."
+)
+
+if __name__ == "__main__":
+    demo.launch()