Shubh Saravgi

Software Engineer with 3.8 years of experience, including 2 years focused on Generative AI, building and deploying internal AI systems for telecom log analysis and workflow automation. Hands-on experience across the full GenAI stack — RAG pipelines, LangChain/LangGraph agentic workflows, semantic search, transformer-based classification, and automated report generation. Skilled in building scalable backend services that process large-scale unstructured log data using Python, FastAPI, Django, Docker, and AWS.

Software Engineer — GenAI / Backend at Samsung R&D India (2022-06 – Present)

• Fine-tuned and deployed transformer-based models (BERT, DistilBERT) for multi-label classification of complex telecom network logs, achieving over 92% accuracy on real internal telecom network event data.
• Designed and optimized ETL pipelines using pandas and regex to transform large-scale unstructured telecom logs into structured datasets, enabling downstream NLP and LLM processing.
• Built end-to-end RAG pipelines using sentence-transformers for embedding generation and ChromaDB for vector storage and semantic retrieval, improving retrieval accuracy by 30% over keyword-based baselines.
• Developed NLP and transformer-based summarization pipelines using spaCy and LLMs to surface actionable insights from noisy, sequential telecom log data.
• Integrated internal GPT-class LLM APIs into team workflows for text generation, structured output extraction, summarization, and automated report generation — including response validation and fallback handling.
• Built LangChain and LangGraph-based agentic workflows integrating tool calling, external APIs, and stateful context management for multi-step automated report generation on telecom log data.
• Delivered low-latency inference APIs using FastAPI and async I/O, actively used by 25+ engineers within the team; served via Django WSGI servers (Gunicorn/uWSGI).
• Architected and launched a Technical Report Generation Platform integrating NLP pipelines, LLMs, and RAG workflows — reducing manual report preparation effort by 90% and saving engineers approximately 45 minutes per report.

B.Tech in Computer Science & Engineering – University of Petroleum and Energy Studies (2018 – 2022)