Projects and Case Studies | Bikash Sapkota

Mobility Data Foundation: GPS-scale geospatial pipelines

Problem: Raw mobility inputs are noisy, high-volume, and difficult to reuse across research, consulting, analytics, and product workflows.

Solution: Designed reproducible ingestion and transformation layers that validate, normalize, aggregate, enrich, and prepare location records for downstream analytics.

Impact: Reduced repeated preprocessing work by standardizing analytical datasets. Improved confidence in outputs through validation, documentation, and repeatable transformations. Created foundations for maps, dashboards, CSV exports, and model features.

Stack: Python, Spark, AWS S3, AWS Glue, GeoPandas, H3, Data quality

People-Flow Analytics Products: Urban planning and location intelligence

Problem: Urban and mobility teams need reliable views of movement patterns without manually interpreting raw GPS trajectories.

Solution: Built analytical outputs for OD movement, road-volume intensity, and POI-footfall patterns using spatial aggregation and map-ready data products.

Impact: Translated movement signals into decision-ready geospatial outputs. Supported stakeholder workflows across research, consulting, and product teams. Made repeated mobility analysis easier to reproduce and compare over time.

Stack: Python, H3, GeoPandas, ClickHouse, Maps, CSV products

Energy Pricing Optimization System: Optimization and platform integration

Problem: Pricing workflows needed optimization logic that could be integrated into operational systems and maintained over time.

Solution: Implemented optimization modules, API interfaces, and OpenADR-related integrations while improving architecture reliability.

Impact: Moved optimization research closer to production workflows. Created integration surfaces for client platforms. Reduced operational risk by refactoring away single points of failure.

Stack: Python, MINLP, REST APIs, OpenADR, Optimization, System refactoring

Document Intelligence Pipeline: OCR and operational automation

Problem: Scanned claims required structured extraction and classification before they could move efficiently through operational workflows.

Solution: Combined OCR extraction, text classification, rule-based extraction, NER, and interface improvements to support faster manual review.

Impact: Improved the path from scanned documents to structured operational data. Reduced friction for manual keying workflows. Connected ML extraction with practical back-office usability.

Stack: OCR, Tesseract, FineReader, WEKA, Random Forest, NER