Just Dial — Internal ML Product
Designed and deployed LSTM-based deep learning models to automatically audit tele-marketing call quality — replacing a fully manual review process and scaling to 3 lakh+ minutes of audio per month with near-zero human effort.
Just Dial operates one of India's largest tele-marketing operations, generating hundreds of thousands of minutes of call recordings monthly. Quality audit was entirely manual — a team of reviewers listened to random samples, an inherently slow and unscalable process that covered only a fraction of total calls.
The goal was to build an automated ML pipeline that could process every call, classify quality parameters, detect compliance issues, and flag problematic calls — all without human involvement.
Custom audio preprocessing pipeline extracting MFCCs, spectral features, and temporal patterns from raw call recordings.
Sequence-to-label LSTM models trained on labeled call data to classify call quality, compliance, and script adherence.
Dashboard showing audit results, flagged calls, trend analysis, and per-agent performance metrics in real time.
Automatic flagging of calls that deviate significantly from expected patterns — catching compliance violations early.
Built an end-to-end pipeline from raw call recordings to processed features, handling audio normalization, segmentation, and feature extraction at scale.
Trained LSTM models in PyTorch on labeled call data, iterating on architecture and hyperparameters to achieve 95%+ accuracy on quality classification.
Deployed as a containerized service processing call recordings in batches, with results feeding into the audit dashboard and alerting systems.
I build deep learning pipelines from data processing to production deployment.