United States
Software Engineer at Siemens with a Master’s in NLP & Machine Learning from Carnegie Mellon University. I specialize in building AI-powered workflows for Siemens’s flagship CAD software: NX, applying expertise in LLMs, NLP, and deep learning to deliver scalable solutions impacting millions of users. My background combines academic rigor in AI research (text generation, speech, machine translation) with industry experience deploying cloud-based and open-source LLMs into production across FinTech, engineering, and biotech.
- Developed a Temporally grounded Video Retrieval Augmented Generation (RAG) chatbot for Siemens NX in a highly Agile development team. - Improved Vector database (ChromaDB) MAP by 5 points by switching from SentenceTransformers embeddings to the OpenAI text-ada-002 embedding model. - Augmented retrieval index with QA pairs and video scene descriptions generated using Claude-3-Sonnet and Claude-3-Haiku (AWS Bedrock). - Identified 3 most common RAG failure cases using RAGAS evaluation suite and GPT-4 (Azure OpenAI). - Optimized GPT-4 inference (number of Azure calls and tokens per minute) by 40% using a combination of streaming, dynamic batching, and generation size control.
10601 - Introduction to Machine Learning - Designed and graded homework assignments and exams for a class of 400 students under Prof. Matt Gormley and Prof. Henry Chai. - Improved weekly office hour attendance by 50% by mentoring undergraduate and graduate students. - Led a team of TAs to create exam questions on ML topics (logistic regression, reinforcement learning algorithms).
- Worked with Prof. Carolyn Rose on improving extended text generation for medical use cases. - Proposed a novel architecture for turn-wise text generation using a cascade of Llama2-7b-chat-hf generator and Maximal Marginal Relevance (MMR) summarizer models. - Performed prompt engineering with parameteric medical persona blog posts to evaluate turn-wise text consistency. - Analyzed the generations (blogs) qualitatively, and quantitatively by proposing 5 new factual plausibility metrics.
Teaching Assistant for CS F241 Microprocessor Programming & Interfacing.
- Augmented customer complaints platform database with 6 months of external data scraped using Reddit API. - Improved Reddit thread analysis efficiency by 30% by optimizing the unsupervized intent detection codebase. - Enhanced existing Reddit pipeline by implementing an unsupervised aspect-based sentiment analysis module (sentenceBERT, nltk.vader) and a novel retrieval metric for ’engagement’. - Consolidated the Reddit insights pipeline into single internal tool for CFR, through cross-functional collaboration.