Open to opportunities

Devansh Vora

@devanshvora

Message

I build scalable LLM and knowledge-graph systems for real-world reasoning.

India

Message

What I'm looking for

I’m looking for a team where I can build LLM + knowledge-graph systems, run rigorous evaluations, and ship production-ready pipelines—especially with agentic orchestration and RAG—while growing fast through research-to-deployment ownership.

I’m an AI research professional focused on building scalable systems for knowledge-graph construction, multi-hop reasoning, and modern LLM applications. Currently, I’m a Jr. AI Research Fellow exploring agentic pipelines that extract structure and meaning from unstructured text.

In my work on SAGE-KG, I architected a schema-free, domain-agnostic agentic knowledge graph framework that decomposes extraction into Fact Extractor, E-R Schema Planner, and Triplet Generator roles. With explicit state transitions and controlled inter-agent information flow, I achieved 92.95% factual capture on MINE and 95% entity recovery on SemEval-2010, outperforming GPT-4o-based KGGen baselines.

I also developed evaluation-driven integrations across HotpotQA, MuSiQue, and 2WikiMultiHopQA, and demonstrated scalability by combining SAGE-KG with HippoRAG2 for 64.4% Exact Match on 2WikiMultiHopQA—while reducing LLM call overhead by 73% versus KGGen under constrained resources.

Earlier, I contributed as an AI Research Intern by fine-tuning Vision Transformers/Swin Transformers for brand infringement detection and deploying a Docker-based RunPod serverless endpoint. I’ve also built GenAI applications with RAG (including legal terminology/IPC), and delivered backend capabilities with Django REST and FastAPI.

Experience

Work history, roles, and key accomplishments

Current

Jr. AI Research Fellow

Current

Smart Energy Learning Center - Dhirubhai Ambani University

May 2025 - Present (1 year)

Architected SAGE-KG, a schema-free agentic knowledge-graph construction framework that achieved 92.95% factual capture on MINE and 95% entity recovery on SemEval-2010 without fine-tuning or predefined ontologies. Built a sequential agent orchestration pipeline and, when integrated with HippoRAG2, improved 2WikiMultiHopQA to 64.4% Exact Match while reducing LLM call overhead by 73%.

Orchestration Knowledge Graphs Retrieval Augmented Generation (RAG)PyTorch Transformers Docker AWS Sagemaker Evaluation

AI Research Intern

Communication Crafts Pvt. Ltd.

Dec 2024 - May 2025 (5 months)

Finetuned Vision Transformers and Swin Transformers for brand infringement detection using FP16 mixed precision, introducing a custom vectorized loss function that improved accuracy 3x over traditional contrastive loss. Deployed a serverless RunPod GPU endpoint with Docker using batch processing and implemented Flash Window Attention to preserve pretrained Swin weights.

Swin Transformer Vision Transformer PyTorch Flash Attention Docker RunPod Deployment Computer Vision Custom Loss Function

AI Engineer Intern

Cilans Systems Pvt. Ltd.

Jun 2024 - Jul 2024 (1 month)

Developed a GenAI application that analyzed agent–customer transcripts using LLMs (Mistral-Large, GPT-4) to automatically generate audit reports on response quality, empathy, and problem-solving. Implemented legal terminology RAG for Indian IPC sections using GPT-4, Mistral, Gemini, and Llama3 with Hugging Face, Ollama, and LangChain; used AWS Bedrock for Mistral inference.

LLM Mistral GPT 4 Retrieval Augmented Generation (RAG)LLMs (LangChain Hugging Face Ollama Prompt Engineering

Django Backend Intern

DevNullX Technology Pvt. Ltd.

Dec 2023 - Feb 2024 (2 months)

Built a Salesman Expense and Data Tracking system in Python-Django to compute expenses using MapMyIndia axial distance based on entered journey coordinates. Visualized sales routes and locations on an interactive map with Folium to improve route management and expense monitoring.

Python Django Geospatial Distance Computation MapMyIndia Data Processing Analytics