🚀 FINAL POLISH: Fixed image visibility in UI and added captioned plots to README

README.md

@@ -152,13 +152,25 @@ LLM Agent (GRPO trained)
 
 ![Score Heatmap](assets/score_heatmap.png)
 
-*Each column is one episode. Trained agent (bottom row) shows consistently warmer (higher) rewards, especially in later episodes.*
+*Figure 2: Heatmap showing per-episode rewards. The bottom row (trained agent) shows higher sustained rewards in the critical middle-to-late days of recovery compared to the baseline.*
 
 ### Reward Curve Over Training
 
 ![Training Loss](assets/training_loss.png)
 
-*4-episode moving average. Trained agent steadily improves above greedy baseline.*
+*Figure 3: 4-episode moving average. The Sarvam-105B agent steadily learns to capture both service restoration and fairness bonuses, outperforming the heuristic greedy baseline after ~20 iterations.*
+
+### Utility vs Fairness Trade-off
+
+![Utility vs Fairness](assets/utility_vs_fairness.png)
+
+*Figure 4: Intersectional analysis showing the agent's progress. Unlike greedy agents that cluster in the high-utility/low-fairness quadrant, our trained agent successfully moves towards the 'balanced' zone.*
+
+### Fairness Progress
+
+![Fairness Improvement](assets/fairness_vs_episode.png)
+
+*Figure 5: Total Fairness Score across episodes. The training successfully pushed the agent to consider vulnerable zones, resulting in a consistent upward trend in equity achievement.*
 
 ### Key Numbers
 

server/app.py

@@ -105,12 +105,20 @@ def _build_app():
 
         with gr.Tab("Analysis & Fairness"):
             gr.Markdown("## 📊 Training Results & Fairness Trends")
+            
+            # Use absolute filesystem paths for gr.Image
+            base_dir = os.path.dirname(os.path.dirname(__file__))
+            results_img = os.path.join(base_dir, "assets", "training_results.png")
+            heatmap_img = os.path.join(base_dir, "assets", "score_heatmap.png")
+            loss_img = os.path.join(base_dir, "assets", "training_loss.png")
+            fair_img = os.path.join(base_dir, "assets", "fairness_vs_episode.png")
+
             with gr.Row():
-                gr.Image("/assets/training_results.png", label="Trained vs Baseline")
-                gr.Image("/assets/score_heatmap.png", label="Episode Rewards")
+                gr.Image(results_img if os.path.exists(results_img) else None, label="Trained vs Baseline")
+                gr.Image(heatmap_img if os.path.exists(heatmap_img) else None, label="Episode Rewards")
             with gr.Row():
-                gr.Image("/assets/training_loss.png", label="Reward Convergence")
-                gr.Image("/assets/fairness_vs_episode.png", label="Fairness Improvement")
+                gr.Image(loss_img if os.path.exists(loss_img) else None, label="Reward Convergence")
+                gr.Image(fair_img if os.path.exists(fair_img) else None, label="Fairness Improvement")
 
         with gr.Tab("README"):
             readme_path = os.path.join(os.path.dirname(os.path.dirname(__file__)), "README.md")