Varun Billuri

Built supervised fine-tuning (SFT) pipelines for small language models (SLMs), covering multilingual dataset preparation, training, evaluation, and deployment. Evaluated fine-tuned models against baselines and benchmarked LLM, STT, and TTS providers on multilingual accuracy, response quality, and inference latency to guide model selection. Developed an Arabic dialect evaluation platform integrating LLM prompting, context engineering, and speech workflows. Built real-time healthcare conversational AI integrating translation, speech recognition, and neural speech synthesis for Tigrinya, Wolof, Bambara, and Arabic dialects. Redesigned production AI pipelines to improve inference scalability, multilingual orchestration, and cross-service reliability. Applied reinforcement learning from human feedback (RLHF) to align multilingual model responses with human preferences.

Developed an ML-driven incident response pipeline that routed Slack alerts and created Jira tickets, reducing mean time to resolution (MTTR) by approximately 60%. Built an anomaly detection service for Airflow execution metadata, surfacing pipeline failures and bottlenecks through PostgreSQL-backed dashboards. Engineered an agentic AI framework using LangChain and LangGraph to triage Airflow failures through rule-based logic, LLM classification, and guarded automated remediation. Deployed LLM-based log summarization with LangChain and Ollama, converting raw Airflow logs into actionable debugging insights. Extended petabyte-scale ETL orchestration across AWS and GCP, modularizing data pipelines that supplied ML training datasets. Contributed to supervised fine-tuning of Azure OpenAI models for internal workflows, including dataset preprocessing and model evaluation.

Developed a retrieval-augmented generation (RAG) prototype using LangChain and Haystack, indexing pricing and metrics data in vector stores for contextual question answering. Experimented with embedding models to improve retrieval relevance and support contextual responses. Built backend services and data pipelines for pricing and metrics datasets, enabling downstream machine learning workflows. Prototyped AI workflows combining retrieval, summarization, and alerting to automate recurring analysis tasks.