Designed and built a production-grade LLM-powered internal platform that transforms crossteam

communication into structured project insights, improving visibility across engineering and product

organizations in Meta

• Engineered end-to-end LLM application pipelines integrating LLaMA 3.1 (405B) and LLaMA 3.3 (70B) to

summarize large-scale internal project discussions with minimal manual intervention.

• Developed robust prompt engineering and orchestration logic to extract actionable signals (blockers,

milestones, decisions, risks, strategic highlights) from unstructured, multi-author communication threads.

• Optimized model inference behavior through parameter tuning (temperature, top-p, context windowing)

and prompt iteration to balance factual accuracy, completeness, and response latency.

• Implemented a hash-based in-memory caching layer for partial LLM outputs, reducing redundant inference

calls and significantly improving system throughput and end-to-end latency.

• Built scalable backend APIs exposing AI-powered summarization and insight extraction as reusable services,

enabling integration with internal messaging and project-tracking systems.

• Conducted user-in-the-loop evaluation and iteration cycles, incorporating real user feedback to improve

output reliability and inform broader internal LLM adoption strategies.

• Collaborated cross-functionally with engineers, product managers, and stakeholders

Worked as team lead, designed the scope and timeline. Designed and implemented a regionalized Contact Service that stores customers contacts, provides APIs for CRUD operations. The application also include an asynchronous workflow that publish data to the Centralized DataAnalysis platform. Enabled the merchant to see aggregated business insights in comparison to order or product level data, merchants can customize clustered information by their behaviors and habits, thus resulting in better business predictions, increase sales and reduce costs.

• Leveraged API Gateway Lambda authorizer to implement AWS auth with resource based access control.

• Utilized Top Level Resources for  fine-grained access control and data isolation that ensures data security.

• Designed and implemented an ECS Fargate service that provides APIs to manipulate customer contact data. Used DDB encryption client to encrypt personal data in DDB Table and created an Index table to handle different query use cases.

• Configured DynamoDB stream, Lambda function, SNS, SQS with Pub/Sub pattern to fanout data into Index table and publish to EventBus that flow into the Centralized Data Analysis platform.

• Designed and implemented exponential backoff  retry strategy in synchronized API and error handling for async workflow. Created metrics and alarms on failed messages and documented DLQ re-drive SOP for operational excellence.

• Developed script to handle customer DSAR and Data deletion requests to compliant with GDPR and CCPA.

Core Services in Multi-Channel Fufillment

• Based on new customer requirements, add interactions to the SIRM(ShipmentInjectionRequestManager ) orchestrators, which orchestrates a minimum of 6 services to validate and produce a valid order.

• Optimized auth solution in SIRM to extend transitive auth and  fine-grained access control to Amazon at large guarantee customer trust in their data privacy.

• Build services for new public APIs using Java with new service infrastructure using AWS API gateway, NLB, ECS fargate. And use DDB stream, Lambda function, SQS and SNS for asynchronous processes.

• Used the Gizmo throttling to apply API level access rates from different resources, which helps ensure availability of host resources, maintain optimal resource utilization, and increase confidence in SLAs promised to your customers.

Built a data pipeline that migrated legacy site data for more than 400 existing warehouses of the East Region Team(managing areas of NY, NJ and so on) from local storage to AWS Redshift

• Leveraged infrastructure as code, Typescript CDK(Cloud Development Kit) , created monitors and alarms for the automation system. Enabled automatic backup and restore. Resulted in simplfied automating operational work.

• Conducted design reviews for AWS automated data normalization work flow with senior staff  members in the team.

• Ran integration tests, unit tests, canary test and e2e (end to end) tests, and successfully launched. Individually designed various AWS resources(Lambda, S3, and Step Functions etc) in yaml format to simulate actual computing and storage scenarios for TDD (Test Driven Development). Leveraged Python Openpyxl to implement Lambdas business logic of DRS interaction.

• Created EventBridge triggers that allowed Redshift to import new data automatically and a life cycle management for S3 to expire uploaded data.

• Anually tested various scenarios in AWS CLI to enable elective problem-solving like race conditions handling.

Drove discussions for the long term solution of Web Based DRS UI. Selected the best practice among QuickBase, QuickSight, and React-Redux built from scratch.

• Designed a warehouse Data Management systems to fully automate Design Engineers' retrospective data review process which reduced more than 20 Design Engineer working hours to 2 seconds.

• Created a database management system on Redshift, which shall be consumed by the

SCMS(Station Capacity Management System). Enabled client teams (Capacity Planning Team and so on) to query data through Hubble.

• Regularly communicated with customers and QuickBase, the third party developers to ensure 100% coverage of use cases.

• Provided technical guidance and support to customers in implementing and configuring the software solution, ensuring seamless integration with their existing data calculation and data stream work flows.

• Conducted regular meetings and feedback sessions with customers to validate software functionalities related to data calculation and data stream, incorporating their input into iterative development cycles.

Leveraged the component-based architecture with React to create efficient and scalable user interfaces.

• Designed and implemented a workow that leveraged multiple stacks to capture, upload, and compute images and

translate them using S3, Lambda and Step Functions, Amazon Rekognition, and Amazon Translate

• Designed and implemented a notication system to send costumer daily study reports with CloudWatch Event.

Optimized the AWS DynamoDB table by changing indexes and reduced the latency by 40%.

• Implemented UI widgets including log-in, photo capture, of the translation application on Android and IOS mobile

app.