Summary
AI researcher with experience at Amazon AWS and Nokia Bell Labs, focused on building and deploying agentic, tool-augmented LLM systems that deliver measurable impact across production and research settings.
Education
PhD in Computer Engineering, Stony Brook University (GPA: 4.0)
- Research Focus: Neuro-symbolic and tool-augmented large language models for reliable reasoning, with an emphasis on verification-aware prompting, reinforcement learning for tool-use policies, and agentic collaboration in open-ended coding and problem-solving tasks.
MS in Computer Engineering, Stony Brook University (GPA: 3.8)
Skills
| Programming | Python, C++, JavaScript, SQL |
| LLMs & Reasoning | Tool-Augmented & Agentic Systems, Prompt Optimization (APO), Function Calling, Reinforcement Learning, Chain-of-Thought, Verification-Aware Reasoning |
| Model Training & Adaptation | Fine-tuning (LoRA, QLoRA), Instruction Tuning, Evaluation & Benchmarking |
| Frameworks & Libraries | PyTorch, TensorFlow, Hugging Face, LangChain, NNsight, SpaCy, NLTK |
| Multimodal & Speech | Automatic Speech Recognition (Whisper), Multimodal Pipelines (Text–Audio–Code) |
| MLOps & Deployment | Git, Docker, Linux, MLflow, DVC, CML, SageMaker, Bedrock |
| Data & Knowledge Systems | Vector Databases (Qdrant, Pinecone), Knowledge Graphs, Neo4j, GraphRAG |
Publications
2025
- Shaik, H., Villuri, G., & Doboli, A. (2025). Two Large Language Model-based Methods to Validate Open-Ended Problem Solving in Teams.
- Shaik, H., Villuri, G., & Doboli, A. (2025). An Overview of LLMs and a Novel, LLM-Based Cognitive Architecture for Solving Open-Ended Problems. (MAKE 2025)
- Shaik, H., Villuri, G., & Doboli, A. (2025). Concept Combinations with Generator and Validator Agents Prompted Using Insights from Concept Networks. (Complex Networks 2025)
- Villuri, G., & Doboli, A. (2025). An Experimental Study on the Interpretability of Transformer Models for Dialog Understanding. (CAI 2025)
- Villuri, G., Shaik, H., & Doboli, A. (2025). A Stacked Multi-Layered Perceptron - LLM Model for Extracting the Relations in Textual Descriptions. (SSCI 2025)
- Villuri, G., Doboli, A., & Pallapu, H. R. (2025). Towards Semantic Classification: An Experimental Study on Automated Understanding of the Meaning of Verbal Utterances.
2024
- Villuri, G., & Doboli, A. (2024). An Overview and Discussion of the Suitability of Existing Speech Datasets to Train Machine Learning Models for Collective Problem Solving. (arXiv)
- Villuri, G., & Doboli, A. (2024). Using Speech Data to Automatically Characterize Team Effectiveness to Optimize Power Distribution in Internet-of-Things Applications.
- Pallapu, H. R., Villuri, G., Doboli, A., & Doboli, S. (2024). Automatically Understanding Human Behavior for IoT Applications with Optimized Human-in-the-Loop Control.
- Villuri, G., & Doboli, A. (2024). Towards Semantic Classification of Dialog using Contextual Prediction Networks. (CCN 2024)
Research Experience
Graduate Research Assistant – Stony Brook University
- Built LLM-based validation agents to detect errors in open-ended collaborative coding tasks.
- Implemented verification-aware generator–validator pipelines for static and real-time team evaluation.
- Designed neuro-symbolic dialogue and code representations using operational and axiomatic semantics.
- Developed tool-augmented LLM pipelines with knowledge graphs and multimodal NLP.
- Created graph-augmented agent frameworks for relational reasoning and agentic collaboration.
- Applied reinforcement learning and multi-agent methods for tool-use and coordination policies.
- Led research on reliable LLM reasoning by integrating validation, symbolic structure, and agentic workflows, resulting in peer-reviewed publications across multiple venues.
Professional Experience
Applied Scientist Intern – Amazon AWS (Mountain View, CA)
- Evaluated the “GEPA” Automatic Prompt Optimization (APO) framework for replacing hand-crafted prompts with automated optimization strategies.
- Validated structured reasoning accuracy improvements up to 12% via algorithmic search over instruction space.
- Benchmarked optimization across 10 production LLMs (1B–405B), identifying a “sweet spot” where mid-sized models yielded highest relative gains.
- Studied cross-model transferability: prompts optimized for mid-sized models generalized well to larger architectures, enabling prompt reuse.
ML Research Intern – Nokia Bell Labs (Murray Hill, NJ)
- Developed a multi-agent LLM pipeline using OPC UA to extract insights from industrial telemetry data.
- Designed hierarchical agents for high-level question generation and low-level insight extraction.
- Integrated knowledge graphs and function calling to correlate multi-layered industrial events.
- Investigated prompting strategies to enhance LLM accuracy in complex industrial NLP tasks.
Software Developer Intern – Zippi Delivery (Stony Brook, NY)
- Built AI-driven restaurant management platform using GoHighLevel MCP and LLMs for unified business operations.
- Developed multi-agent “Zippi” system with CrewAI achieving 95% task routing accuracy.
- Designed a hybrid LLM setup (Gemini + DeepSeek) cutting costs by 40% and manual oversight by 65%.
- Created webhook automation handling 10K+ social interactions daily, improving response times by 50%.
Projects
SmartTutor: GraphRAG-based Learning Assistant
- Developed an AI-powered tutor leveraging GraphRAG, replacing traditional RAG with a knowledge graph for improved accuracy of 34% and enhanced interpretability.
- Processed 300 pages on Stereo Vision, generating a structured knowledge graph instead of word embeddings.
- Achieved a 2-second response time and cut user study time by 25%, simplifying complex concepts.
- Enabled real-time cache visualization, offering a transparent and explainable retrieval mechanism.
- Deployed a functional prototype using distilGPT2 with an optimized graph-based retriever.
Language Neutralization System
- Architected real-time translation system for call centers, supporting 10+ languages with 95% accuracy.
- Constructed streaming pipeline (ASR → translation → TTS), reducing processing delay by 40%.
- Streamlined system for accelerated machines, yielding 30% performance boost.
- Orchestrated system deployment via Docker, minimizing setup time by 60%.
- Established API web socket, amplifying system accessibility by 80%.
Agent Performance Platform
- Led platform development using text classification to analyze agent conversations, processing 1,000+ daily interactions.
- Fine-tuned model to attain 85% accuracy in identifying and scoring 15 distinct soft skill traits.
- Executed MLOps practices, accelerating model update time by 70% and enhancing overall accuracy by 10%.
- Boosted efficiency by 50% and reduced manual review time by 75%.
Open Source Contributions
-
GEPA:
Configurable
DefaultAdapterevaluation returning per-example score + feedback (PR #147).