diff --git a/backend/app/main.py b/backend/app/main.py
index c1ff323..0fb041c 100644
--- a/backend/app/main.py
+++ b/backend/app/main.py
@@ -266,9 +266,20 @@ def backfill_embeddings():
     threading.Thread(target=_run, daemon=True).start()
 
 
+def _acquire_singleton_lock() -> bool:
+    """Returns True if this worker wins the startup lock for singleton tasks (scheduler, backfill).
+    Uses Redis SETNX with a 5-minute TTL. Degrades gracefully if Redis is unavailable."""
+    try:
+        import redis as redis_lib
+        r = redis_lib.from_url(settings.REDIS_URL, decode_responses=True, socket_connect_timeout=2)
+        return bool(r.set("startup:singleton_lock", "1", nx=True, ex=300))
+    except Exception:
+        return True  # Redis unavailable — assume single worker, run everything
+
+
 @asynccontextmanager
 async def lifespan(app: FastAPI):
-    # Startup
+    # Startup — all workers run DDL/seeds (idempotent), only one runs scheduler/backfill
     Base.metadata.create_all(bind=engine)
     setup_pgvector()
     os.makedirs(settings.UPLOAD_DIR, exist_ok=True)
@@ -276,11 +287,15 @@ async def lifespan(app: FastAPI):
     os.makedirs(settings.CHROMA_PERSIST_DIR, exist_ok=True)
     seed_admin()
     seed_default_models()
-    backfill_embeddings()
-    start_scheduler()
+    # Scheduler and backfill must only run in one worker to avoid duplicate jobs / race conditions
+    is_primary = _acquire_singleton_lock()
+    if is_primary:
+        backfill_embeddings()
+        start_scheduler()
     yield
     # Shutdown
-    stop_scheduler()
+    if is_primary:
+        stop_scheduler()
 
 
 app = FastAPI(
diff --git a/backend/app/routers/admin.py b/backend/app/routers/admin.py
index 82b31a0..fb269b7 100644
--- a/backend/app/routers/admin.py
+++ b/backend/app/routers/admin.py
@@ -142,7 +142,7 @@ def search_litellm_models(
             return {"models": models, "source": base}
         except Exception as e:
             log.warning(f"LiteLLM model search failed: {e}")
-            raise HTTPException(status_code=400, detail="Failed to query models API. Check the API base URL and key.")
+            raise HTTPException(status_code=400, detail=f"Failed to query models API: {e}")
 
     try:
         import litellm
@@ -304,7 +304,7 @@ def search_tts_voices(
             raise
         except Exception as e:
             log.warning(f"ElevenLabs voice discovery failed: {e}")
-            raise HTTPException(status_code=400, detail="ElevenLabs API error — check your API key and try again.")
+            raise HTTPException(status_code=400, detail=f"ElevenLabs API error: {e}")
 
     elif provider == "polly":
         access_key = api_key or settings.AWS_ACCESS_KEY_ID
@@ -337,7 +337,7 @@ def search_tts_voices(
             raise
         except Exception as e:
             log.warning(f"AWS Polly voice discovery failed: {e}")
-            raise HTTPException(status_code=400, detail="AWS Polly error — check your credentials and region.")
+            raise HTTPException(status_code=400, detail=f"AWS Polly error: {e}")
 
     elif provider == "openai":
         voices = []
diff --git a/backend/app/routers/teach.py b/backend/app/routers/teach.py
index e9c99de..7d14aa9 100644
--- a/backend/app/routers/teach.py
+++ b/backend/app/routers/teach.py
@@ -105,7 +105,12 @@ def _build_system_prompt(question: Question, similar: list[Question]) -> str:
     prompt += (
         "\nAnswer the student's question directly. Explain the correct answer, "
         "the underlying concept, and why wrong options are incorrect if relevant. "
-        "Do not ask what they want to know — just teach."
+        "Do not ask what they want to know — just teach.\n\n"
+        "After your explanation, suggest exactly 3 follow-up questions the student might want to ask next. "
+        "Put them at the very end, each on its own line starting with '> '. Example:\n"
+        "> Why is dopamine not the first-line treatment here?\n"
+        "> What are the other causes of this presentation?\n"
+        "> How does this differ in neonates vs older children?"
     )
     return prompt
 
@@ -177,8 +182,18 @@ async def chat(
         if settings.LITELLM_API_BASE:
             kwargs["api_base"] = settings.LITELLM_API_BASE
         response = await litellm.acompletion(**kwargs)
-        reply = response.choices[0].message.content.strip()
-        return {"reply": reply}
+        raw = response.choices[0].message.content.strip()
+
+        # Parse out follow-up suggestions (lines starting with "> ")
+        lines = raw.splitlines()
+        suggestions = [l[2:].strip() for l in lines if l.startswith("> ")]
+        reply_lines = [l for l in lines if not l.startswith("> ")]
+        # Trim trailing blank lines from reply
+        while reply_lines and not reply_lines[-1].strip():
+            reply_lines.pop()
+        reply = "\n".join(reply_lines).strip()
+
+        return {"reply": reply, "suggestions": suggestions[:3]}
     except Exception as e:
         import logging
         logging.getLogger(__name__).error(f"TeachChat error for user {current_user.id} model {model_id}: {e}")
diff --git a/docker-compose.yml b/docker-compose.yml
index 29e5baa..c703d48 100644
--- a/docker-compose.yml
+++ b/docker-compose.yml
@@ -26,6 +26,7 @@ services:
 
   backend:
     build: ./backend
+    command: uvicorn app.main:app --host 0.0.0.0 --port 8000 --workers 4
     env_file:
       - ./backend/.env
     environment:
diff --git a/frontend/src/components/TeachChat.jsx b/frontend/src/components/TeachChat.jsx
index a383093..524e07a 100644
--- a/frontend/src/components/TeachChat.jsx
+++ b/frontend/src/components/TeachChat.jsx
@@ -14,7 +14,7 @@ import api from '../api/client'
  */
 export default function TeachChat({ question }) {
   const [open, setOpen] = useState(false)
-  const [messages, setMessages] = useState([])   // {role, content}
+  const [messages, setMessages] = useState([])   // {role, content, suggestions?}
   const [input, setInput] = useState('')
   const [loading, setLoading] = useState(false)
   const [noModel, setNoModel] = useState(false)
@@ -65,7 +65,7 @@ export default function TeachChat({ question }) {
         messages: next,
         model_id: selectedModelId || null,
       })
-      setMessages(m => [...m, { role: 'assistant', content: res.data.reply }])
+      setMessages(m => [...m, { role: 'assistant', content: res.data.reply, suggestions: res.data.suggestions || [] }])
     } catch (err) {
       const detail = err.response?.data?.detail || 'AI error'
       if (err.response?.status === 503) setNoModel(true)
@@ -187,28 +187,49 @@ export default function TeachChat({ question }) {
               </div>
             )}
             {messages.map((m, i) => (
-              <div key={i} style={{
-                alignSelf: m.role === 'user' ? 'flex-end' : 'flex-start',
-                maxWidth: '85%',
-                background: m.role === 'user' ? 'var(--primary)' : 'var(--bg)',
-                color: m.role === 'user' ? 'white' : 'var(--text)',
-                padding: '8px 12px',
-                borderRadius: m.role === 'user' ? '12px 12px 2px 12px' : '12px 12px 12px 2px',
-                fontSize: '0.85rem',
-                lineHeight: 1.55,
-                border: m.role === 'assistant' ? '1px solid var(--border)' : 'none',
-              }}>
-                {m.role === 'assistant'
-                  ? <ReactMarkdown components={{
-                      p: ({children}) => <p style={{margin: '0 0 6px'}}>{children}</p>,
-                      ul: ({children}) => <ul style={{margin: '4px 0', paddingLeft: 18}}>{children}</ul>,
-                      ol: ({children}) => <ol style={{margin: '4px 0', paddingLeft: 18}}>{children}</ol>,
-                      li: ({children}) => <li style={{marginBottom: 2}}>{children}</li>,
-                      strong: ({children}) => <strong style={{fontWeight: 700}}>{children}</strong>,
-                      code: ({children}) => <code style={{background: 'var(--border)', borderRadius: 3, padding: '1px 4px', fontSize: '0.82em'}}>{children}</code>,
-                    }}>{m.content}</ReactMarkdown>
-                  : m.content
-                }
+              <div key={i} style={{ alignSelf: m.role === 'user' ? 'flex-end' : 'flex-start', maxWidth: '90%', display: 'flex', flexDirection: 'column', gap: 6 }}>
+                <div style={{
+                  background: m.role === 'user' ? 'var(--primary)' : 'var(--bg)',
+                  color: m.role === 'user' ? 'white' : 'var(--text)',
+                  padding: '8px 12px',
+                  borderRadius: m.role === 'user' ? '12px 12px 2px 12px' : '12px 12px 12px 2px',
+                  fontSize: '0.85rem',
+                  lineHeight: 1.55,
+                  border: m.role === 'assistant' ? '1px solid var(--border)' : 'none',
+                }}>
+                  {m.role === 'assistant'
+                    ? <ReactMarkdown components={{
+                        p: ({children}) => <p style={{margin: '0 0 6px'}}>{children}</p>,
+                        ul: ({children}) => <ul style={{margin: '4px 0', paddingLeft: 18}}>{children}</ul>,
+                        ol: ({children}) => <ol style={{margin: '4px 0', paddingLeft: 18}}>{children}</ol>,
+                        li: ({children}) => <li style={{marginBottom: 2}}>{children}</li>,
+                        strong: ({children}) => <strong style={{fontWeight: 700}}>{children}</strong>,
+                        code: ({children}) => <code style={{background: 'var(--border)', borderRadius: 3, padding: '1px 4px', fontSize: '0.82em'}}>{children}</code>,
+                      }}>{m.content}</ReactMarkdown>
+                    : m.content
+                  }
+                </div>
+                {/* Clickable follow-up suggestion chips — only on last assistant message */}
+                {m.role === 'assistant' && m.suggestions?.length > 0 && i === messages.length - 1 && !loading && (
+                  <div style={{ display: 'flex', flexDirection: 'column', gap: 4 }}>
+                    {m.suggestions.map((s, si) => (
+                      <button
+                        key={si}
+                        onClick={() => { setInput(s); setTimeout(() => inputRef.current?.focus(), 50) }}
+                        style={{
+                          textAlign: 'left', background: 'var(--card-bg)', border: '1px solid var(--primary)',
+                          borderRadius: 8, padding: '5px 10px', fontSize: '0.78rem', cursor: 'pointer',
+                          color: 'var(--primary)', lineHeight: 1.4,
+                          transition: 'background 0.15s',
+                        }}
+                        onMouseEnter={e => e.currentTarget.style.background = 'var(--primary-light, rgba(99,102,241,0.08))'}
+                        onMouseLeave={e => e.currentTarget.style.background = 'var(--card-bg)'}
+                      >
+                        ↩ {s}
+                      </button>
+                    ))}
+                  </div>
+                )}
               </div>
             ))}
             {loading && (