Piyush Patil

- Engineered LSTM-based demand forecasting models in PyTorch across 80K+ daily ride requests, tracking experiments with Weights & Biases to systematically improve accuracy during peak-demand periods. - Orchestrated end-to-end ML pipelines using PySpark, Databricks, Delta Lake, and Docker to ingest and transform 35GB+ weekly ride and telemetry data, increasing forecast refresh frequency and reducing retraining overhead. - Formulated surge pricing prediction models using XGBoost, advanced feature engineering, and hyperparameter tuning to estimate fare-adjustment patterns across 3 metropolitan regions. - Constructed geospatial demand prediction models with H3 indexing and spatial feature engineering to pinpoint high-demand zones, cutting rider wait times by 12% in targeted areas. - Versioned 120K+ driver feedback records with DVC and Great Expectations for data validation, producing consistent, schema-validated training datasets for downstream NLP workflows. - Fine-tuned a transformer-based language model for driver feedback classification via QLoRA (PEFT), processing 18K+ monthly records and cutting manual review volume by 300+ cases per month. - Architected an agentic trip-ranking system with LangGraph, vector embeddings, and similarity search to surface personalized ride options for 25K+ weekly users.

- Surfaced 2,150 anomalous patterns across 9 enterprise engagements by training risk prediction models with Scikit-learn, logistic regression, and statistical modeling on 420K+ transactional records. - Established data quality standards by implementing validation frameworks using dbt, Great Expectations, and rule-based PySpark checks across 35+ large-scale datasets during quarterly audit cycles. - Improved high-risk entity identification by deploying gradient boosting and decision tree classification models on 180K+ structured records, delivering measurable gains in precision and recall. - Eliminated 8 hours of manual effort per reporting cycle by re-engineering ETL pipelines with Airflow, dbt, and Pandas, consolidating 62GB of data from 7 disparate source systems. - Strengthened audit coverage via Isolation Forest anomaly detection and statistical outlier analysis across 275K+ records, flagging 3,480 irregular patterns for investigative review. - Accelerated review turnaround by 2 business days by shipping Tableau analytics dashboards and KPI reporting frameworks across 14 cross-functional projects.

- Interrogated 180K+ client and transaction records with SQL and data aggregation techniques, uncovering 1,100+ reporting inconsistencies and improving monthly business reporting accuracy. - Launched Power BI dashboards and DAX reports for business operations and client KPIs, producing 10+ dashboards and reducing manual reporting effort by 5 hours per week. - Audited 15+ datasets through Excel VBA, data reconciliation, and control checks, resolving 800+ mismatched records to improve downstream reporting reliability. - Streamlined data extraction and transformation with Python, Pandas, and ETL pipelines, processing 30GB+ structured business data and trimming weekly preparation time by 6 hours. - Generated trend analysis reports with Excel forecasting models and time-series techniques across 24+ months of production data to support planning and performance tracking. - Applied linear regression and feature selection via Scikit-learn on 90K+ transaction records to expose cost and revenue variance patterns and support evidence-based business decisions.