Building, evaluating and securing AI systems.
I build AI systems with a focus on reliability, evaluation, and real-world failure modes. My work spans machine learning, deep learning, LLMs, RAG, fine-tuning, and agentic AI, with cybersecurity and networking as a specialization.
I care about more than making a model or demo work once. I like to measure it, find where it fails, understand why, and verify whether a fix actually holds up.
Machine Learning & Deep Learning
Model development, graph ML, speech and audio ML, feature engineering, and evaluation.
LLMs & Generative AI
RAG, hybrid retrieval, reranking, QLoRA/DPO fine-tuning, local LLM inference, and agentic AI.
AI Evaluation & Reliability
Benchmark design, retrieval evaluation, faithfulness testing, error analysis, ablation studies, and failure-mode analysis.
AI Security
Prompt injection, adversarial testing, agent and MCP security, groundedness verification, and defensive evaluation.
AI Engineering
Python, PyTorch, FastAPI, Docker, PostgreSQL, Qdrant, REST APIs, Linux, and CI/CD.
Security-advisory extraction using Qwen2.5-3B and QLoRA, with hand-verified data, in-distribution and out-of-distribution evaluation, FastAPI serving, and a code-level groundedness layer for detecting unsupported model claims.
Evaluation-first RAG pipeline over 85 FastAPI documentation files and 756 chunks, comparing dense retrieval, BM25, hybrid retrieval, reranking, and generation faithfulness using an 80-pair verified QA set.
MCP-based AI agent security testbed covering prompt injection, tool-level attacks, adversarial testing, and defense evaluation.
Telugu speech data acquisition and processing pipeline covering licensed data collection, VAD segmentation, and preparation of audio datasets for downstream ASR and ML work.
Security-hardened distributed key-value store inspired by Dynamo, covering replication, quorum operations, vector clocks, mTLS, authentication, encryption, observability, and adversarial testing.
LLM evaluation, agentic AI, multimodal learning, AI security, production ML systems, and MLOps.