Built a generalized file processing system leveraging vLLM, capable of extracting structured data from highly unstructured clinical documents with 99.99%+ accuracy, using a flexible marker-file schema.

Designed an innovative marker-based learning framework to drastically reduce parsing cost and latency while improving extraction accuracy, setting new internal standards for document understanding.

Developed an internal agentic coding assistant that automated repetitive workflows, saving over 80% of engineering time and significantly accelerating customer onboarding.

Designed a multimodal agent framework capable of interpreting on-screen context, user intent, and personalized workflows to deliver seamless, adaptive assistance.

Developed the web console using Next.js and the main application using Electron, integrating both into a cohesive user experience.

Leveraged LiveKit as the core voice infrastructure, and developed a custom in-house AI agent system optimized for latency, reliability, and dynamic task routing.

Personally conducted 50+ client interviews within the first month to gather feedback, validate use cases, and accelerate the development cycle through rapid iteration.

Utilized AWS s3, AWS Textract, AWS CloudFront, AWS LightSail, and MySQL database to construct complete video processing and evaluation pipelines.

Used OOP and Multi-threading to improve the performance of data pipelines.

Individually developed a video ranking system that displays the most appropriate content to the end users.

Individually Constructed an administration panel web application using ant Design Pro, Typescript, and Flask to manage all application users' data.

The application was developed using Amazon's ASK API and deployed on Amazon's EC2 instances for echo devices serving a user base of >50 million, I designed the agents backend using a microservice architecture. Implementing containerization (GPU Docker), caching, load balancing(ECS) for efficient performance and auto-scaling.

Designed and implemented a hybrid system to align LLM-based responses with AWS safety standards, reducing conversation friction and enhancing user response rates by 30%.

Analysed log data to identify & fix bugs with backend services, communicated findings to engineers at Amazon