Hi — I'm Kashish, a CS + Applied Math student at Vanderbilt and a software engineer focused on agentic AI: systems where a model plans, calls the right tools, and stays grounded in real data instead of guessing. Right now I'm a software engineering intern at ID.me, where I architected Member Assistant — a conversational AI agent that orchestrates tools across three products from a single Spring Boot service, built on Google's Agent Development Kit.
Most of my work lives where machine learning meets solid backend engineering: agent orchestration, retrieval pipelines, and the infrastructure that keeps them fast and honest. Before software I competed in extemporaneous debate and co-authored a financial-literacy curriculum for undocumented and unaccompanied minors that scaled across New Jersey — I still care a lot about who a piece of technology is actually for.
Lately I'm going deep on agent frameworks, tool-use reliability, and how you actually evaluate agents. Off the clock: bouldering, film photography, modular synths, and brain–computer interfaces.
Architected Member Assistant, a platform-agnostic conversational AI agent serving RAG-grounded answers across three ID.me products from one Java/Spring Boot microservice on Google's ADK. Built a modular tool-abstraction layer decoupling orchestration from data sources (cut p95 latency ~35%) and the first store-search + community-scoped offer retrieval over Elasticsearch and PostgreSQL.
Re-architected the backend for a learning platform serving 3,800+ users: 15+ REST endpoints, a PostgreSQL migration saving $10K/yr, and a real-time attendance tracker wired into the Zoom API.
Designed and deployed a RAG pipeline (LangChain + Phi) over 100+ company docs, from ingestion and embedding through a production API — cutting support tickets 30% at 92% accuracy.
A production conversational AI agent that grounds every answer in real data — plans and calls tools across Vertex AI Search, Elasticsearch, and PostgreSQL, retrieves over 10K+ offers and 500+ stores, and returns correct per-member cashback at 94% answer relevance.
Explore the interactive breakdown →A multilingual (EN/ES) financial-coaching app with a generative-AI core, helping 50+ students navigate budgeting, banking, and ID acquisition; moving toward a licensable NJ high-school curriculum.
A local, judge-based evaluation pipeline for RAG/agent systems: a synthetic corpus, a ground-truth Q&A set, and an LLM-as-judge scorer validated with Cohen's kappa, all on local inference.
A deadline aggregator unifying Gmail, Calendar, Slack, and Canvas into one view.