• Making DNA Sequencing faster ⚡️

• Deployed a FastAPI inference endpoint to generate embeddings from DNA sequences using a 300M parameter model; integrated with the AIDO Agent to support real-time queries, handling over 1,000 requests per day.

• Optimized API deployment system by implementing a proactive model preloading mechanism, caching high-demand models into CPU memory at server startup, which significantly reduced model load latency and improved request processing speed.

• Built a Kubernetes KFPv2 research pipeline for finetuning LLMs over multi-node GPUs using PyTorch and Wandb reducing experimentation cycle time by 80%. Hosted it using a FastAPI endpoint.

Near Real-Time Issuer Updates:

• Designed a microservice using Spring Boot to manage near real-time card updates from 9000+ issuer banks adding $20M of revenue to $160M previously.

• Spearheaded the development of the File Arrival module, a crucial component within our system architecture.

• Devised solutions to effectively handle files arriving out of order and post cut-off time, enhancing system reliability and coherence.

PAN Delivery Module using Kafka:

• Implemented a delivery module in Kafka to push PAN chains to acquirer banks in real-time;

• Achieved 10x faster writes compared to a MySQL-based process, ensuring scalability on large-scale data loads.

• Enabled a retry mechanism for failed updates ensuring EOS.

AAA-VAU MySQL Profile Data Sync for Daily Batch Processing:

• Developed a multi-threaded, modularized, and config-driven Java utility for importing data from MySQL to Hadoop via Apache Sqoop, utilizing Amazon S3 storage as a drop zone to ensure data consistency in an Active-Active dual datacenter setup. The utility is highly customizable where the pull type can be easily changed by just adding a class through the factory method and a line change in the config.

• Achievement: Reduced data pull time from 80 to 20 minutes and achieved 100% data consistency on both data centers.

Active/Expired PAN Report:

• Engineered a Hive-based solution in Python that efficiently generates a monthly report subscribed to by 75+ issuer banks globally.

• A card is Active/Expired when marked as "Active" on our system but has actually "Expired". This occurs when issuer banks send incorrect expiry dates.

• Optimized card filtering process by joining data from multiple Hive tables, managing a substantial volume of approximately 89 million cards.

• Leveraged Tez to expedite processing, improving operational efficiency and reducing data retrieval latency.

NOTE: Ensured the integrity and reliability of all code through CI/CD pipelines incorporating SonarQube, Sonatype, Checkmarx, and Blackduck scans to maintain high standards of code quality and security.

• Developed a Management Decision Support System for the Traveling Salesman Problem (MDSS-TSP) in Python, optimizing the order of outlet assignment to minimize travel costs for salesmen as a part of Operations Research.

• Employed and visualized methods including Enumeration, Branch & Bound, Hungarian Assignment, and Nearest Neighbor in Python.

• Conducted thorough analysis on multiple 10x10 distance matrices, comparing algorithmic time complexity and efficiency to provide actionable insights.

• Used CPLEX libraries to evaluate algorithmic time complexity; uncovered optimization opportunities leading to a 40% reduction in processing time for critical operations.