I am a Generative AI Engineer with 2.5 years of production experience architecting enterprise-grade LLM applications, multimodal AI pipelines, and high-performance inference platforms. My work spans fine-tuning foundation models (QLoRA/SFT), optimizing latency and compute costs via AutoRound W4A16 quantization and vLLM multi-GPU serving, and building autonomous agentic workflows.

• Current Focus: Fine-tuning VLMs/LLMs, accelerating high-throughput vLLM serving, and optimizing Graph-RAG architectures.
• Impact Highlight: Creator of a quantized model on Hugging Face with 100K+ downloads.
• Core Philosophy: "Optimization is not just about speed; it's about making advanced AI architecturally sustainable."

• Fine-Tuning: Domain adaptation using QLoRA, SFT, Unsloth, and Hugging Face PEFT.
• Quantization: W4A16/INT4 model compression using AutoRound, AWQ, and GPTQ.
• High-Throughput Serving: Production multi-GPU deployment via vLLM, cutting P95 latency by 50%.

• Hybrid Retrieval: Orchestrating BM25, semantic vector search, and Reciprocal Rank Fusion (RRF) in Qdrant and Neo4j.
• Graph-RAG Architectures: Implementing LightRAG and property graph strategies using Qdrant and Neo4j for dual-level knowledge extraction.
• Observability & Evals: Production tracing, prompt evaluation, and hallucination tracking with Langfuse and Ragas.

• Stateful Agents: Designing autonomous multi-agent loops, self-healing tool execution, and session memory using LangGraph.
• Multimodal Pipelines: Real-time object detection (YOLOv12) and speech-to-text processing (Whisper ASR) for automated verification platforms.

• Agentic LightRAG for Executive Audit Review: Built a LangGraph multi-agent loop with LightRAG, Qdrant and Neo4j acting as a document, vector, and property graph store for long-form audit analysis.
• Agentic Hybrid RAG for Regulatory Compliance: Engineered a self-healing RAG platform in Qdrant with BM25 + Vector RRF search, intercepting DB exceptions for auto-correction and 99.9% uptime.
• AI-Driven vKYC Automation System: Built an end-to-end multimodal pipeline featuring YOLOv12 object detection, Whisper ASR, and FastAPI to streamline document and audio verification.

• Published the gemma-4-E4B-it-W4A16-AutoRound-GPTQ quantized model, surpassing 100K+ downloads.

• Contributed official integrations for the Google Gemini API and PostgreSQL multi-tenant workspace isolation.
• PRs: #2538 | #2556 | #2615

• Identified critical W4A16 serialization bugs in AWQ export for Qwen3-VL models to improve vLLM inference stability.
• Issue: intel/auto-round #1377

• LinkedIn: vishva-r
• Instagram: @justt_vishva
• Portfolio: GitHub Repositories

⭐️ Building the infrastructure that makes AI smarter, faster, and more accessible.