Intelligent Call Center Routing

A closed-loop ML pipeline that routes customer calls using personality-aware NLP, reinforcement learning, and deep outcome prediction — with a feedback engine that learns from every interaction.

Active Research
System Architecture

Three models, one feedback loop

Each call flows through three ML layers — NLP intake, RL routing, and outcome prediction — then feeds back into the system. Models learn from each other, from outcomes, and from human feedback.

🔍
NLP Intake
RoBERTa (125M params)
Classifies intent, sentiment, urgency, complexity, and caller personality from real transcripts. Five fine-tuned heads on a shared backbone.
🔀
RL Router
PPO (Stable-Baselines3)
Learns to assign calls to advisors by optimizing a multi-objective reward: resolution quality, efficiency, satisfaction, workload balance, and burnout protection.
📊
Outcome Predictor
LSTM Multi-Head (PyTorch)
Predicts handle time, first-call resolution, and customer satisfaction using advisor history sequences. Serves as the router's world model.
Key Innovation

Personality-driven data generation

Most call center models treat callers as interchangeable data points. CEPM assigns each caller a personality profile that drives every downstream metric — not just labels, but features the models learn from.

Jung's 12 Archetypes

WHO the caller is

Innocent, Orphan, Hero, Caregiver, Explorer, Rebel, Lover, Creator, Jester, Sage, Magician, Ruler. Each defines baseline behavioral traits: patience, volatility, assertiveness, technical literacy.

Campbell's Hero's Journey

WHERE they are emotionally

Ordinary World, Call to Adventure, Refusal, Meeting the Mentor, Crossing the Threshold, Tests, Ordeal, Reward, Road Back, Return. Each phase modifies the base personality based on emotional state.

120 explicit profiles (12 × 10) produce measurably different call outcomes:

Rebel × Ordeal
Escalation modifier
3.24x
Sage × Ordeal
Escalation modifier
1.08x
Innocent × Ordinary
Escalation modifier
0.30x
Scale
92K
Real transcripts
120
Archetype profiles
3
ML paradigms
83+
Automated tests
Built With
PyTorch HuggingFace Transformers Stable-Baselines3 Gymnasium MLflow FastAPI Next.js pandas ONNX