AI/ML & Data Engineer | MS in Computer Science @ Arizona State University (4.0 GPA)

I build not only intelligent, low-latency scalable AI systems at the intersection of large language models, multi-agent RAG, and large-scale data engineering but also, implement continuous operational visibility, performance diagnostics, drift detection, and safety governance for LLM applications. Previously worked ~5 years at Dell Technologies as Senior Data Engineer, engineering real-time fraud-detection pipelines and language-agnostic semantic retrieval systems that led to a granted patent in hybrid search.

• 🔭 Building systems using LangGraph, GraphRAG, and vLLM optimization pipelines with AI Observability & Governance
• 🎓 MS CS @ ASU (Aug 2025 – May 2027) | B.Tech+M.Tech @ IIIT Gwalior
• 📄 Patent holder — Distributed Hybrid Search for Language-Agnostic Retrieval (US 2024)
• 📫 Reach me at daripa.ushnesha9701@gmail.com

AI / ML & LLMs

Data & Systems

Databases & Cloud

📈 Robust Financial Multi-Agent GraphRAG & Guardrail System

• Engineered a LangGraph multi-agent pipeline (Supervisor, Context Retriever, Web Searcher, Python Math Executor) to orchestrate financial reasoning over the 6K+ FinQA Benchmark Dataset.
• Built a hybrid search engine integrating Neo4j (GraphRAG with 20K+ nodes), Qdrant (dense vectors), and BM25 (lexical search) with a Cross-Encoder Reranker.
• Implemented an LLM-as-a-judge evaluation suite, and optimized performance to achieve a 65% reduction in TTFT (under 120ms) and an 82% prefix cache hit rate using vLLM chunked prefill and Arize Phoenix tracing.

🛰️ AI-Powered Autonomous Trash Interceptor

• Developed a real-time object tracking system on a Raspberry Pi using an ArduCAM ToF depth sensor, processing frames at 30 FPS to detect flying projectiles via background subtraction.
• Implemented a dual trajectory prediction pipeline pairing a Kalman filter physics model with a RandomForest model trained on 500+ throws to predict landing spots within 150ms.
• Engineered a closed-loop PID controller to drive a 4-motor Mecanum wheel platform, achieving projectile intercept catch rates of 85%+ on physical hardware.

🐦 BirdCLEF+ 2026 (Kaggle Competition)

• Pretrained and fine-tuned EfficientNet variations to identify 234 wildlife species from audio recordings of Brazil's Pantanal wetlands, achieving a score of 0.85.
• Converted raw audio into mel-spectrograms, trained with data augmentation and AUC-optimized loss, and ensembled predictions for 5-second segments of one-minute soundscapes.