Parth Bansal

As a Software Engineering Intern at this Agri-Tech startup, I focus on building robust data pipelines and machine learning solutions to drive spatial and agricultural insights. Key contributions include: Spatial Data Engineering: Developed and optimized a PostGIS ETL pipeline covering 733 districts, leveraging GIST indexing for high-performance spatial queries. Large-Scale Data Processing: Validated and reconciled massive agricultural datasets, including over 173,000 CCE records and 90,000+ APY entries, to successfully resolve critical LGD code and product mismatches. Machine Learning Solutions: Designed a highly available KNN-based yield estimator equipped with multi-tiered fallbacks to effectively handle sparse data. Backend API Development: Architected a unified API surface that delivers actionable planned-project insights - such as nearby project counts, growth metrics, and top builder scores - which directly support urban valuation and banker reports.

I engineered the computer vision (CV) backend for a mobile-first AI image editor as part of the Inter IIT Tech Meet 14.0. My focus was on building a fast, highly realistic, and context-aware editing pipeline. Key technical highlights include: Advanced Contextual Compositing: Architected a "Draft-Critique-Refine" pipeline utilizing InternVL to critique scene lighting, geometry, and physics. This guided Flux.1 to resolve floating subjects and shadows, resulting in artifact-free, highly realistic image compositing. High-Speed Semantic Edits: Designed semantic color grading via Mask2Former for localized OpenCV edits, achieving an ultra-low latency of just 0.9 seconds. Precision Style Control: Developed a structure-aware workflow using ControlNet and IP-Adapter to lock original image geometry while seamlessly injecting new styles. 2D to 3D Reconstruction: Built a feature leveraging Large Gaussian Models to convert standard 2D inputs into fully textured 3D meshes. Inference Optimization: Optimized backend performance by implementing a Singleton Model Cache, which significantly cut the cold-start overhead for diffusion backbones.