Hello, I'm Om Sethi

• Re-engineered a proof-of-concept into a React and Node.js MVP using IBM Consulting Advantage to generate synthetic PDFs, emails, and spreadsheets for model training
• Evaluated and benchmarked an AI content-generation platform against reference documents, systematically testing single and multi-reference model configurations to surface accuracy, formatting, and hallucination failure modes
• Debugged and root-caused recurring defects in a document-generation pipeline, including mapping errors, hierarchy loss, and missing sections, then fed findings back into template and mapping configuration to improve output reliability

• Architected an LLM document pipeline on AWS Bedrock using a three-tier model dispatcher that routes 95% of calls to Haiku, escalates to Sonnet on low confidence, and applies Claude vision OCR with per-citation source verification
• Built a HIPAA and SOC 2-oriented Next.js and React monorepo on AWS ECS Fargate with blue/green CI/CD, Aurora Postgres Serverless v2 row-level security, and in-house argon2id authentication
• Shipped to production at a paying law firm, powering AI splitting of 1,000+ page discovery PDFs, automated medical index generation, and citation-backed Statement of Facts drafting for attorney review

• Maintaining and scaling an academic TraceTool used in CSE 115 and 116, enabling students to trace code execution and visualize memory state, supporting 500+ students annually and 5,000+ submissions per semester
• Onboarding into a large existing codebase while contributing to system scalability efforts focused on multi-institution deployment, higher concurrency, and reliable handling of peak submission loads
• Working with course staff to ensure correct and consistent memory modeling and tracing behavior, improving reliability and clarity of the tool as student usage continues to grow

• Built scalable backend systems with FastAPI and Docker for AI-powered applications
• Optimized machine learning workflows using LangChain and Ollama for efficient model deployment
• Developed secure, reliable multi-agent routing systems for complex AI task execution
• Implemented prompt-based task execution frameworks for enhanced AI interaction capabilities

A real-time hand gesture control system that enables volume control, play-pause, app switching with 95% gesture recognition accuracy and under 200ms response time.

• Currently developing a real-time hand gesture control system using Python, MediaPipe, and OpenCV with 95% gesture recognition accuracy and under 200ms response time
• Building a modular gesture recognition architecture currently supporting 4+ gesture types (thumbs up/down, fist, palm) using object-oriented design patterns achieving 25-30 FPS real time processing
• Implementing a computer vision pipeline that processes live video streaming with 21-point hand landmark detection, featuring two-phase execution system (200ms activation delay + 600ms cooldown) to prevent false triggers

A Discord music bot that streams high-quality audio, supports playlists and queue management and offers interactive commands for seamless group listening.

• Built high-quality audio streaming with support for YouTube
• Implemented playlist management with load functionality and queue persistence
• Created interactive commands for play, pause, and skip
• Developed user-friendly interface with real-time status updates
• Added support for multiple voice channels and server-specific configurations