Harsh Mehta is an AI Engineer in Charlotte, NC with 5+ years across data science and machine learning, including 2+ years shipping production LLM systems. He currently builds RAG and LangGraph agent systems at Kaiser Permanente.

What does Harsh Mehta do at Kaiser Permanente?

He builds citation-backed RAG over 5,000+ pages of compliance documentation, with LangGraph agents and human-in-the-loop review. The work cut a 3-month policy review to one week and reclaimed roughly 300 analyst-hours per cycle.

Does Harsh Mehta have LangChain and LangGraph experience?

Yes. He has shipped LangGraph and LangChain multi-agent systems in production, including agent orchestration at Kaiser Permanente, the multi-agent core of the Ensemble booking platform on Vertex AI, and 7 multi-agent LLM workflows at ATC.

Is Harsh Mehta open to new roles?

Yes. He is open to AI Engineer, Machine Learning Engineer, and Data Scientist roles in Charlotte, NC and remote across the US.

Where did Harsh Mehta go to school?

He holds an M.S. in Information (Data, Machine Learning, Cloud) from the University of Wisconsin-Madison with a 3.95 GPA, and a Bachelor of Management Studies from the University of Mumbai with a 3.5 GPA.

What certifications does Harsh Mehta hold?

He holds the Databricks Data Engineer Associate, Snowflake SnowPro Core, Azure DP-100, Power BI PL-300, and AZ-900 certifications (all 2025), plus Salesforce Tableau.

← Back

Extern · Aug 2025 - Apr 2026 · Remote

AI Engineer (Externship)

Outamation: document intelligence for 200+ page mortgage files with degraded scans and brutal multi-column layouts. Benchmarked 4 OCR engines and 3 embedding strategies on cost, latency, and accuracy, then let the evidence pick the stack: ChandraOCR and DocTR, BGE embeddings, FAISS, DeepSeek-R1 for grounded generation. 95% extraction accuracy with page-level citations, deployed via Replicate and Gradio.

Wayfair: LangChain agents for SKU onboarding and catalog enrichment at 95% task completion, batch and real-time inference on AWS (S3, Lambda, SageMaker) with vector search, redundant LLM API calls cut 35%. Profiled an A100 pipeline everyone assumed needed more hardware and found the GPU idle most of the wall clock: batching and caching took per-document embedding from 120s to 15s, about 8x, on the same machine.

95% extraction accuracy on 200+ page mortgage files with page-level citations
Benchmarked 4 OCR engines and 3 embedding strategies on cost, latency, and accuracy
LangChain agents at 95% task completion for SKU onboarding and catalog enrichment
A100 per-document embedding 120s to 15s (~8x) on the same GPU

OCRFAISSAWSLangChain