Call Copilot

Confidence score
Objection Handling
Pitch

Inspiration

We built Call Copilot around one principle from the Pillar 01 brief:

“The agent uses your tool. Your tool doesn’t use them.”

Instead of building an auto-dialer or lead prioritizer, we focused on helping agents during active conversations. The goal was simple: help agents understand how to talk to a customer, not who to call next.

What it does

Call Copilot is an AI-powered conversation assistant for bank call-center agents.

Given a customer profile, the system: predicts calibrated subscription likelihood using LightGBM explains the prediction with SHAP generates personalized talking points using Llama 3.3 provides contextual objection handling during live calls

The tool never auto-dials, ranks customers, or makes approval-style decisions. The agent stays fully in control.

How we built it

Stack: Python, pandas, scikit-learn LightGBM, SHAP, Optuna Groq Llama 3.3 70B Streamlit Key ML decisions Removed duration to avoid data leakage Optimized F2 instead of accuracy due to class imbalance Used scale_pos_weight ≈ 7.9 Applied isotonic probability calibration Tuned threshold for recall-sensitive performance Results PR-AUC: 0.470 ROC-AUC: 0.814 F2: 0.589

Challenges we ran into

Avoiding data leakage from post-call features Designing the UI to comply with Pillar 01 rules Balancing PR-AUC vs F2 during tuning Preventing robotic LLM-generated scripts Accomplishments that we’re proud of Honest calibrated probabilities Fully explainable predictions Human-centered AI workflow Real-time contextual objection handling Strong performance on heavily imbalanced data

What we learned

We learned that: calibration builds trust metric selection matters more than raw accuracy explainability is critical in human-facing AI systems constraints lead to better product design

What’s next

Voice transcription with Whisper Drift monitoring and auto-retraining Real-time spoken objection handling Supervisor analytics dashboard

Tagline

Helping call-center agents know what to say — not who to call.