AI Engineer — M.S. Computer Science @ University of Florida
I build LLM-powered systems that go beyond demos — grounded retrieval, multi-agent orchestration, and production deployment on real infrastructure.

Currently building a RAG-based grant writing assistant deployed on Azure at Street Care.
🔍 Open to AI Engineer / ML Engineer roles — graduating Dec 2025.

I don't just fine-tune models. I build the full system around them:

• Retrieval pipelines that actually work — chunking strategy, embedding choice, hybrid search, metadata filtering
• Agentic workflows where agents are independently testable and replaceable
• Evaluation layers — LLM-as-Judge, retrieval benchmarking, grounding checks — because shipping without measurement isn't production
• Deployed APIs — FastAPI, Docker, GitHub Actions CI/CD, Azure Container Apps

Multi-agent orchestration for automated research workflows

Planner → Search → Scrape → Retrieve → Writer → Evaluator

• ~15–20s end-to-end latency across complex multi-hop queries
• 5–8 live sources per query with fully citation-backed outputs
• 50+ query evaluation using LLM-as-Judge across relevance, faithfulness, and completeness
• Each agent independently testable and replaceable — built for production-style modular control → View on GitHub

Grounded Q&A over research-paper corpora

• End-to-end RAG system: PDF ingestion → chunking → MiniLM embeddings → FAISS indexing → FastAPI → Streamlit
• Retrieval relevance improved from 68.5% → 80.0% through iterative tuning of chunk size, overlap, and top-k, plus corpus expansion across RAG, FAISS, BM25, and LoRA papers
• Modular architecture designed to scale toward 10,000+ document chunks
• Built evaluation utilities for corpus statistics, retrieval benchmarking, and grounding checks
• Strict retrieval-grounded prompting to reduce hallucination risk → View on GitHub

End-to-end personalized outreach pipeline

• LLM pipeline: job description → ChromaDB semantic retrieval → personalized cold email
• Modular 3-stage architecture: ingestion → retrieval → generation
• Configurable tone, sender profile, and project context for targeted outreach
• GitHub Actions CI runs on every push for stable, deployment-ready builds → View on GitHub

Production-grade financial risk classification

• P95 latency reduced by 39% — from 2.7ms to 1.6ms — by eliminating redundant preprocessing
• Real-time predictions via FastAPI REST API with end-to-end scikit-learn pipeline
• SMOTE-Tomek for robust class imbalance handling
• Containerized with Docker, CI/CD via GitHub Actions → View on GitHub

• RAG-based grant writing assistant on Azure (Street Care) — recursive chunking, metadata filtering, hybrid search
• Stronger evaluation loops across agentic pipelines
• Production-ready RAG workflows with measurable retrieval quality benchmarks

🌐 Portfolio → tvanish002.github.io
💼 LinkedIn → linkedin.com/in/anish-tv