Hi, I'm Kaushik

Migrated EC2 batch workload to scheduled AWS Lambda — cut runtime from ~12 min to ~4 min while lowering compute costs.

Designed automated rollback & failure-notification workflows, restoring app state on failed executions.

Integrated Splunk for centralized logging & observability across the stack.

Developed features across Spring Boot / Angular and a secondary Python / Flask app.

Co-developed the Spring Boot backend for a fully offline desktop application with a Vue.js + Electron frontend, on a cross-functional team of 5 (2 backend, 2 frontend, 1 DevOps), bundling a Java runtime inside Electron so the entire backend runs locally with no internet dependency.

Designed and built a local ingestion pipeline that parses and indexes ~1 GB engine documentation ZIP packages in ~15 minutes, handling structured extraction and storage of thousands of technical manual pages for offline retrieval.

Implemented SQLite FTS5 full-text search delivering sub-50ms retrieval across ~5,000 documents, enabling technicians to instantly search dense technical manuals without any network connection.

Introduced a Caffeine caching layer and Spring WebMVC-based document streaming to minimize memory overhead and deliver large manual files efficiently in resource-constrained, offline field environments.

Automated cross-platform build and deployment using AWS CodeBuild, producing signed Windows and Linux installers and publishing them to S3 for distribution.

Extended the architecture to support a LAN mode, allowing mobile and tablet devices to connect to a desktop instance over the local network and access manuals without installing the full application.

Architected an end-to-end OCR + parallel LLM inference + deterministic XML pipeline that converts 20-page scanned engine manuals into structured XML in ~6–7 minutes, replacing a process that was previously impractical to perform manually at scale.

Designed an event-driven ingestion flow where PDFs uploaded to S3 automatically trigger a Lambda function, which orchestrates multi-tasked ECS jobs for parallel OCR preprocessing before handing off to downstream staged processing.

Built a two-stage LLM processing layer that generates per-page structured XML using multi-stage few-shot and XML-constrained prompting, then applies a deterministic XML builder that reuses prior page context to maintain consistency across the document.

Developed an automated LLM-assisted repair step to detect and fix cross-page merge inconsistencies, reducing manual reconciliation of OCR and LLM-generated output errors.

Scaled the pipeline horizontally using concurrent AWS ECS tasks, Lambda orchestration, and CloudFormation-defined infrastructure, enabling reliable throughput as document volume increased.

Served as senior developer on a 6-person cross-functional team (2 Python backend engineers, 1 DevOps, 3 QA testers), owning core architecture and pipeline design decisions.