Srikar Veluvali

Autonomous Fault Recovery in DAG Workflows via Dynamic ReAct Orchestration

Built a YAML-driven directed acyclic graph (DAG) runtime with asyncio execution and 22 node types supporting branching, loops, and integrations across OpenAI, Anthropic, and Bedrock. Added a ReAct self-healing subsystem that autonomously diagnoses failures, applies guarded patches, and rolls back unsafe actions across a 199+ test suite, packaged with a FastAPI backend and an interactive React Flow DAG editor.

Privacy-Preserving Clinical Question Answering via Parameter-Efficient Fine-Tuning and Localized RAG

Built a local, privacy-preserving medical QA assistant by fine-tuning Llama 3 8B Instruct on 100k ChatDoctor conversations via rank-16 4-bit QLoRA, quantizing the result into a 4.69 GB GGUF for local Ollama execution. Coupled the model with a Chroma RAG pipeline chunking 21 medical texts to stream grounded answers with verifiable citations through a secure React and Flask interface.

Stratified Cardiovascular Risk Estimation and Clinical Interpretation via Ensembled Classifiers and LLMs

Developed a predictive screening platform comparing multiple ML models on UCI and BRFSS datasets, serving an optimized logistic regression classifier (88.5% accuracy, 0.89 F1) via Flask APIs. Integrated Google Gemini to translate raw clinical probabilities into actionable lifestyle, dietary, and report interpretations alongside Leaflet provider discovery, earning a People's Choice Award nomination in the Google Gemini API Developer Competition.

Intent-Driven Cross-Platform E-Commerce Retrieval via Dense Embeddings and Real-Time Scraping

Built a full-stack shopping assistant that extracts purchase intent via Groq, matches queries to live Amazon and Flipkart listings through Hugging Face embeddings in Pinecone, and scrapes product data using concurrent aiohttp routines with retry fallbacks. Shipped a responsive React and Flask platform featuring normalized pricing, rating comparisons, JWT authentication, and personalized recommendation histories.

Enhancing Mathematical Chain-of-Thought in Compact LLMs via GRPO

Small language models struggle with multi-step mathematical reasoning and structured output without resource-heavy full-parameter fine-tuning. Engineered a reinforcement learning pipeline adapting Qwen2.5-3B-Instruct to GSM8K using Group Relative Policy Optimization (GRPO) combined with rank-64 4-bit QLoRA and custom XML reasoning tags. Trained across 250 GRPO steps using TRL, vLLM, and Unsloth to evaluate candidate generations against schema and numeric exact-match rewards, yielding competitive chain-of-thought accuracy on a single GPU.

Grounded Multimodal Synthesis and Adaptive Learning via Page-Aware Dense Vector Retrieval

Developed an end-to-end learning platform powered by a page-aware RAG pipeline using 768-dimensional all-mpnet-base-v2 embeddings, Pinecone, and Groq inference for grounded document Q&A and automated study aid generation. Scaled the system across 61 FastAPI routes and 15 MongoDB collections to power mock exams, AI interview practice, mind maps, and interactive TipTap rich-text notebooks.