AI / Data
Engineer

AI for Oil Drilling: Architected a hybrid ML system from the ground up, combining classification, entity extraction, and proprietary algorithms to infer drilling operations & parameters and reconstruct activity timelines from unstructured reports with 95%+ accuracy, saving 4 hrs/well across 1000s of wells.

Selected a combination of small LLMs (qwen, gpt-oss, gemini), fast fine-tuned transformer models (BERT, GLiNER), and decision trees to meet strict latency and cost constraints. Implemented context-aware corrections and domain-specific guardrails to ensure physically consistent outputs.

Data mining: Engineered a pipeline to mine structured insights from PDFs using multi-modal LLMs (gemini-2.5-flash) with pre-defined schemas, few shot examples, and text rules; outperformed SOTA tools like Docling for our use case. Optimized w/ annotation caching, input pruning, etc.

Vessel & Fuel Analytics: Designed a data warehousing solution & a Dagster pipeline to process vessel telemetry & geospatial data, saving clients $1.2M annually in fuel costs through clever fuel reporting using PostgreSQL, Dagster, Polars, Power BI.

Leadership: Promoted to technical lead, responsible for architectural decisions and team building to drive automation, platform modernization. Led the delivery of 10+ mission-critical projects with a team of 6 engineers.

Assisted with data modelling and implementing data pipelines to extract, load, and transform raw Facebook Ads data for BI and Analytics teams.

Migrated complex pipelines from Salesforce Datorama and Adverity to a custom in-house solution using GCP BigQuery, dbt, Apache Airflow, and Python.

Assembled an ETL pipeline from Wikidata to the MusicBrainz database, facilitating a 60% increase in new location data, and slashing manual data feeding by 90% [Details].

Independently designed and developed a scalable, production-ready solution using Python (pandas, multiprocessing, requests, sqlalchemy), SQL (PostgreSQL), and Docker [Architecture, Code].

Conducted research and experimentation, optimizing SPARQL queries to cater to Wikidata’s graph data structure to improve data quality and extraction efficiency. [Details].

Enriched, cleaned, and combined 27 billion rows of music streaming data using Python (Pandas, Multiprocessing), SQL (PostgreSQL), and Apache Arrow – achieving high efficiency in Python without Spark. [Details]

Researched and implemented technologies like Zstandard and Apache Arrow to optimize data lake efficiency, resulting in a 53% reduction in storage and a 9% improvement in read/write speeds.[Details]

Performed data analytics and published benchmarks, dashboards, and reports to help collaborating teams better understand and utilize the data to train state-of-the-art Music Recommendation Systems. [Project Summary]