Work

Worked on Oracle’s JSON-to-Duality Migrator—helping turn document collections into clean relational models when you only have metadata, not the underlying data.

• Designed and implemented schema-only inference using LLMs to identify primary keys and merge semantically equivalent entities from JSON metadata, enabling accurate relational modeling without source data.
• Built core components in C and validated mappings with SQL.
• Added semantic guardrails and metadata-driven heuristics for structural correctness—reducing tables by 25%, improving deduplication by 38%, and successfully migrating 99.7% of documents.

Built a research data pipeline to reconstruct the historical footprint of fintech apps for market analysis.

• Reconstructed ~5GB of historical data across 650 fintech apps by combining the Wayback CDX API with proxy scraping, prioritizing completeness and consistency.
• Used NLP and fuzzy matching (~95% accuracy) to standardize and link entity names across sources, producing a high-completion (~90%) time-series dataset for downstream analysis.

Shipped machine learning for cricket video analysis and helped improve the product’s API and mobile experience.

• Built an ML model to identify shot, stance, and posture from cricket batting video using OpenCV, exposed via REST APIs (~86% accuracy).
• Contributed to Django APIs and Flutter client work to improve functionality and user experience.