Abhijay Sai Paladugu

• Implemented sparse attention mechanisms within the Needle framework, focusing on the "fixed" attention pattern that combines local windowed attention with strided global attention.
• Developed both Python-level attention mechanisms with additive masking and custom CUDA kernels for accelerated sparse computation.
• Implemented and compared multiple positional embedding mechanisms (Learned, Sinusoidal, and RoPE) to enable models to handle longer sequences beyond their training length.

• Developed a comprehensive evaluation framework for assessing long-context capabilities in agentic search systems.
• Designed evaluation protocols across four agentic domains (Search, Code, Reasoning, Tool-use) and integrated three synthetic long-context benchmarks into the framework.
• Implemented a Consistency-Accuracy Index (CAI) metric to analyze task correlations and model consistency across different settings and tasks.

Collaborating with Amazon to design and implement an end-to-end multimodal Retrieval-Augmented Generation (mRAG) pipeline, integrating LLMs with specific image and text retrieval engines for diverse visual question answering datasets, thereby enhancing context accuracy and mitigating hallucinations.

Developed a retrieval-augmented generation (RAG) system using Langchain, FAISS and few-shot prompting to enhance the performance of the Mistral-7B model for factual Q&A about Pittsburgh and CMU achieving 61% F1-score with no fine-tuning.

Improved the CodeRAG Bench by using fixed-size and semantic pre-retrieval chunking in the retrieval pipeline, obtaining ~31% increase in the baseline values on the ODEX coding tasks (open domain) dataset.

• Built a scalable movie recommendation system using collaborative filtering with matrix factorization (SVD) and user segmentation.
• Containerized the inference service with Docker and orchestrated using `docker-compose`; automated model retraining via cronjobs.
• Used Prometheus & Grafana for monitoring model accuracy and system health; implemented A/B testing for model evaluation.
• Tracked model lineage and versions with Weights & Biases (W&B) to ensure reproducibility and provenance of predictions.

• Engineered a cloud-native microservice system to process & serve insights over 1TB of Twitter data using EKS, RDS MySQL, and Spark on Azure.
• Achieved 34k RPS under budget constraints via schema optimization, Java Vert.x web-tier, and AWS ARM-based instances with auto-scaled deployment.
• Used Helm, Terraform, and GitHub Actions for end-to-end CI/CD; monitored performance via Prometheus, Grafana, and AWS CloudWatch.
• Placed 5th in live benchmark test (77.5/80) by fine-tuning query pipelines, deploying sidecar authentication, and optimizing container orchestration.

• Built a generative music system using LlamaIndex and a multi-agent framework consisting of Melody, Harmony, Rhythm, Arrangement, and Critique agents.
• Each agent independently handled its musical aspect using natural language reasoning, contributing to the progressive refinement of musical pieces.
• Implemented inter-agent communication using LLM-based agent orchestration and context-passing via dynamic memory modules in LlamaIndex.
• Critique Agent evaluated the musical output and guided iterative improvements based on stylistic coherence and harmonic balance.
• Exported generated melodies into MIDI format for playback and further processing with DAWs like Ableton and LMMS.