diff --git a/app/classifier.py b/app/classifier.py
index cf8581e..b954ab1 100644
--- a/app/classifier.py
+++ b/app/classifier.py
@@ -1,14 +1,18 @@
-# Simple rule-based urgency classifier. Replace with sklearn/transformer later.
-KEYWORDS_HIGH = {"emergency", "urgent", "help now", "asap", "please help", "immediately", "can't", "cant", "cannot", "can't breathe"}
-KEYWORDS_MED = {"issue", "problem", "not working", "error", "fail", "unable"}
+from app.config import settings
+from typing import Tuple
 
+# Wrapper that delegates to rule-based or ML classifier depending on settings.CLASSIFIER_MODE
 
-def classify_urgency(text: str):
-    t = (text or "").lower()
-    for k in KEYWORDS_HIGH:
-        if k in t:
-            return "high", 0.99
-    for k in KEYWORDS_MED:
-        if k in t:
-            return "medium", 0.75
-    return "low", 0.3
+def classify_urgency(text: str) -> Tuple[str, float]:
+    mode = (settings.CLASSIFIER_MODE or "rule").lower()
+    if mode == "ml":
+        try:
+            from app.classifier_ml_wrapper import classify_urgency_ml
+            return classify_urgency_ml(text)
+        except Exception:
+            # fallback to rule-based
+            from app.classifier_rule import classify_urgency_rule
+            return classify_urgency_rule(text)
+    else:
+        from app.classifier_rule import classify_urgency_rule
+        return classify_urgency_rule(text)
diff --git a/app/classifier_ml_wrapper.py b/app/classifier_ml_wrapper.py
new file mode 100644
index 0000000..fa9d0f7
--- /dev/null
+++ b/app/classifier_ml_wrapper.py
@@ -0,0 +1,28 @@
+import os
+import joblib
+from app.config import settings
+from typing import Tuple
+
+# lazy load to avoid startup cost
+_model = None
+
+
+def classify_urgency_ml(text: str) -> Tuple[str, float]:
+    global _model
+    model_path = os.getenv("MODEL_PATH") or settings.MODEL_PATH
+    if _model is None:
+        try:
+            _model = joblib.load(model_path)
+        except Exception:
+            # If model not available, fall back to rule-based
+            from app.classifier_rule import classify_urgency_rule
+            return classify_urgency_rule(text)
+    try:
+        probs = _model.predict_proba([text])[0]
+        idx = probs.argmax()
+        label = _model.classes_[idx]
+        confidence = float(probs[idx])
+        return label, confidence
+    except Exception:
+        from app.classifier_rule import classify_urgency_rule
+        return classify_urgency_rule(text)
diff --git a/app/classifier_rule.py b/app/classifier_rule.py
new file mode 100644
index 0000000..130bcba
--- /dev/null
+++ b/app/classifier_rule.py
@@ -0,0 +1,14 @@
+# Simple rule-based urgency classifier kept as a fallback and default implementation.
+KEYWORDS_HIGH = {"emergency", "urgent", "help now", "asap", "please help", "immediately", "can't", "cant", "cannot", "can't breathe", "911"}
+KEYWORDS_MED = {"issue", "problem", "not working", "error", "fail", "unable", "bug", "crash"}
+
+
+def classify_urgency_rule(text: str):
+    t = (text or "").lower()
+    for k in KEYWORDS_HIGH:
+        if k in t:
+            return "high", 0.99
+    for k in KEYWORDS_MED:
+        if k in t:
+            return "medium", 0.75
+    return "low", 0.3
diff --git a/app/config.py b/app/config.py
index 3030389..39fee21 100644
--- a/app/config.py
+++ b/app/config.py
@@ -9,6 +9,19 @@ class Settings(BaseSettings):
     PORT: int = 8000
     WEBHOOK_SECRET_TOKEN: str = ""
 
+    # ML & model settings
+    CLASSIFIER_MODE: str = "rule"  # 'rule' or 'ml'
+    MODEL_PATH: str = "models/urgency_model.joblib"
+
+    # Messaging provider
+    WHATSAPP_PROVIDER: str = "twilio"
+
+    # Meta settings (optional)
+    META_TOKEN: str = ""
+    META_PHONE_NUMBER_ID: str = ""
+    META_APP_SECRET: str = ""
+    META_VERIFY_TOKEN: str = ""
+
     class Config:
         env_file = ".env"
 
diff --git a/app/ml_classifier.py b/app/ml_classifier.py
new file mode 100644
index 0000000..cfca252
--- /dev/null
+++ b/app/ml_classifier.py
@@ -0,0 +1,23 @@
+import joblib
+from typing import Tuple
+from sklearn.pipeline import Pipeline
+
+_model = None
+
+
+def load_model(path: str) -> Pipeline:
+    global _model
+    if _model is None:
+        _model = joblib.load(path)
+    return _model
+
+
+def predict_text(model: Pipeline, text: str) -> Tuple[str, float]:
+    """Return (label, confidence)"""
+    if not text:
+        return "low", 0.0
+    probs = model.predict_proba([text])[0]
+    idx = probs.argmax()
+    label = model.classes_[idx]
+    confidence = float(probs[idx])
+    return label, confidence
diff --git a/app/reply_generator.py b/app/reply_generator.py
index 15a36d6..393992a 100644
--- a/app/reply_generator.py
+++ b/app/reply_generator.py
@@ -1,25 +1,48 @@
 import os
 import openai
+import asyncio
+from tenacity import retry, wait_exponential, stop_after_attempt, retry_if_exception_type
 from app.config import settings
 
 openai.api_key = settings.OPENAI_API_KEY
 
+SYSTEM_PROMPT = (
+    "You are a concise, empathetic assistant that helps users over WhatsApp. "
+    "When appropriate, ask clarifying questions. When a message appears urgent, suggest escalation. "
+    "Keep replies short (1-3 sentences) and actionable."
+)
+
+
+@retry(wait=wait_exponential(min=1, max=10), stop=stop_after_attempt(3), retry=retry_if_exception_type(Exception))
+def _call_openai_sync(messages, model="gpt-3.5-turbo"):
+    resp = openai.ChatCompletion.create(model=model, messages=messages, max_tokens=200, temperature=0.3)
+    return resp.choices[0].message.content.strip()
+
+
 async def generate_reply(incoming_text: str, history):
-    # If OPENAI_API_KEY is set, use ChatCompletion for contextual replies, otherwise fallback.
-    if settings.OPENAI_API_KEY:
-        # Build a simple conversation for the LLM from history
-        messages = [
-            {"role": "system", "content": "You are a helpful assistant that replies concisely and empathetically."}
-        ]
-        for item in history:
-            # assume message dicts have 'text' and maybe 'sender' (not enforced here)
-            messages.append({"role": "user", "content": item.get("text", "")})
-        messages.append({"role": "user", "content": incoming_text})
-        try:
-            resp = openai.ChatCompletion.create(model="gpt-3.5-turbo", messages=messages, max_tokens=200)
-            return resp.choices[0].message.content.strip()
-        except Exception:
-            # fallback
-            return "Thanks — I received your message. Can you tell me more?"
-    else:
+    # If OPENAI_API_KEY not set, return fallback
+    if not settings.OPENAI_API_KEY:
+        return "Thanks — I received your message. Can you tell me more?"
+
+    # If history is long, compress/summarize the earlier part (simple approach: keep last 8 messages)
+    recent = history[-8:] if history and len(history) > 8 else history
+
+    messages = [
+        {"role": "system", "content": SYSTEM_PROMPT}
+    ]
+
+    # Add history as user messages
+    if recent:
+        for item in recent:
+            text = item.get("text", "")
+            messages.append({"role": "user", "content": text})
+
+    messages.append({"role": "user", "content": incoming_text})
+
+    # Call OpenAI synchronously inside thread to keep compatibility with sync SDK
+    loop = asyncio.get_event_loop()
+    try:
+        reply = await loop.run_in_executor(None, _call_openai_sync, messages)
+        return reply
+    except Exception:
         return "Thanks — I received your message. Can you tell me more?"
diff --git a/app/train_classifier.py b/app/train_classifier.py
new file mode 100644
index 0000000..ff0bcb4
--- /dev/null
+++ b/app/train_classifier.py
@@ -0,0 +1,39 @@
+import argparse
+import pandas as pd
+from sklearn.linear_model import LogisticRegression
+from sklearn.feature_extraction.text import TfidfVectorizer
+from sklearn.pipeline import Pipeline
+from sklearn.model_selection import train_test_split
+from sklearn.metrics import classification_report
+import joblib
+
+
+def train(input_csv: str, output_path: str, test_size: float = 0.2, random_state: int = 42):
+    df = pd.read_csv(input_csv)
+    if "text" not in df.columns or "label" not in df.columns:
+        raise SystemExit("Input CSV must contain 'text' and 'label' columns")
+    X = df["text"].astype(str)
+    y = df["label"].astype(str)
+
+    X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=test_size, random_state=random_state)
+
+    pipeline = Pipeline([
+        ("tfidf", TfidfVectorizer(ngram_range=(1,2), max_features=10000)),
+        ("clf", LogisticRegression(max_iter=1000, class_weight='balanced'))
+    ])
+
+    pipeline.fit(X_train, y_train)
+
+    preds = pipeline.predict(X_test)
+    print(classification_report(y_test, preds))
+
+    joblib.dump(pipeline, output_path)
+    print(f"Saved model to {output_path}")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="Train urgency classifier")
+    parser.add_argument("--input", required=True, help="Path to labeled CSV (columns: text,label)")
+    parser.add_argument("--output", required=True, help="Output path for model joblib")
+    args = parser.parse_args()
+    train(args.input, args.output)
diff --git a/data/sample_labeled.csv b/data/sample_labeled.csv
new file mode 100644
index 0000000..fae0d05
--- /dev/null
+++ b/data/sample_labeled.csv
@@ -0,0 +1,10 @@
+text,label
+"I need help now, my server is down",high
+"Please respond asap, it's urgent",high
+"I can't login and need access",high
+"I'm seeing an error when I try to login",medium
+"The API returns 500 sometimes",medium
+"Facing intermittent failures in the app",medium
+"What are your pricing plans?",low
+"How do I upgrade?",low
+"Hello, I have a general question",low
diff --git a/docs/ML_AND_OPENAI.md b/docs/ML_AND_OPENAI.md
new file mode 100644
index 0000000..e98001c
--- /dev/null
+++ b/docs/ML_AND_OPENAI.md
@@ -0,0 +1,22 @@
+# ML & OpenAI usage
+
+This document explains how to train the TF-IDF + LogisticRegression urgency classifier and how to enable OpenAI replies.
+
+Training the classifier
+1. Install requirements: pip install -r requirements.txt
+2. Train using the sample data or your own labeled CSV (columns: text,label):
+   python app/train_classifier.py --input data/sample_labeled.csv --output models/urgency_model.joblib
+3. The output model will be saved to models/urgency_model.joblib. Commit it if you want it in the repo (optional).
+
+Using the ML classifier at runtime
+- Set environment variable to use the ML classifier:
+  export CLASSIFIER_MODE=ml
+- Ensure MODEL_PATH (or settings.MODEL_PATH) points to the trained model file.
+
+OpenAI reply generator
+- Set OPENAI_API_KEY in your .env file or environment variables.
+- The reply generator will call OpenAI ChatCompletion (gpt-3.5-turbo by default). If the key is missing or OpenAI fails, the system falls back to a short canned reply.
+
+Notes
+- The app defaults to the rule-based classifier if CLASSIFIER_MODE is not set to 'ml' or if the model fails to load.
+- Keep model files out of source control for larger models; store in an artifacts bucket or volume for production.
diff --git a/requirements.txt b/requirements.txt
index 29de474..84003b7 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -6,3 +6,7 @@ twilio
 openai
 pydantic
 pytest
+scikit-learn
+joblib
+pandas
+tenacity
diff --git a/tests/test_ml_classifier.py b/tests/test_ml_classifier.py
new file mode 100644
index 0000000..3905e0d
--- /dev/null
+++ b/tests/test_ml_classifier.py
@@ -0,0 +1,18 @@
+import pytest
+from app.train_classifier import train
+from sklearn.pipeline import Pipeline
+import joblib
+import os
+
+
+def test_train_and_predict(tmp_path):
+    # train on sample data and ensure model saves and can predict
+    input_csv = os.path.join(os.getcwd(), "data", "sample_labeled.csv")
+    output = tmp_path / "model.joblib"
+    train(input_csv, str(output))
+    assert output.exists()
+    model = joblib.load(str(output))
+    assert isinstance(model, Pipeline)
+    # quick prediction checks
+    pred = model.predict(["This is urgent, please help now"])[0]
+    assert pred in {"high", "medium", "low"}