Looking for a job

Avijit Shil

@avijitshil

AI researcher and engineer advancing generative and agentic AI; built Q-VEX, a compressed Graph-Vector database with sub-25ms hybrid retrieval.

India

Message

What I'm looking for

I’m focused on advancing generative and agentic AI systems, particularly in computer vision, LLM evaluation, knowledge graphs, and multimodal architectures.

With Q-VEX, I architected a local-first Graph-Vector database that combines vector quantization, SQLite FTS5 search, and graph traversal. It achieved 14.5× storage compression and sub-25ms hybrid retrieval, with adapters for LangGraph and CrewAI.

At Scientific Analysis Group (SAG), Defence Research & Development Organisation (DRDO), I formulated a diagnostic framework for object detector failures under black-box patch attacks. The work distinguishes target erasure from semantic class substitution and is under review at SAG Lab, DRDO, MoD.

My research also includes SKG-Eval, a knowledge-graph-based framework for evaluating multi-turn dialogue, and projects in multimodal medical AI and agentic voice systems. I’m pursuing a B.Tech in Electronics and Communication Engineering at Dr. B. C. Roy Engineering College.

Experience

Work history, roles, and key accomplishments

Internship

National Institute of Technology, Durgapur

Worked on an AI Voice Agent for personalized health and nutrition guidance under the mentorship of Dr. Suman Samui.

AI Voice Agents Health Nutrition

Lead Researcher

Independent Project

Architected an interpretable dialogue evaluation framework modeling multi-turn conversations as evolving Semantic Knowledge Graphs (SKGs).

Knowledge Graphs LLM Evaluation Dialogue Evaluation

Collaboration

Massachusetts Institute of Technology (MIT)

Collaborating on a novel mathematical and machine learning framework for robust entity resolution across heterogeneous cross-warehouse databases.

Entity Resolution Embedding models Machine Learning

Internship

Scientific Analysis Group (SAG), DRDO

Formulated an empirical diagnostic framework for object detector breakdown under black-box patch attacks, establishing a taxonomy distinguishing target erasure from semantic class substitution.

Object Detection Black Box Attacks Computer Vision

Lead Researcher

National Institute of Technology (NIT), Durgapur

Formulating an independent verification framework to mitigate cross-modal hallucination loops in forensic Video Question Answering.

Video QA Hallucination Mitigation Multi Modal Models

Collaboration

IIT Jammu & University of Toronto

Developed a segmentation pipeline for shoulder scout MRI images using state-of-the-art models including nnU-Net and MedSAM, collaborating with researchers from IIT Jammu and the University of Toronto.

Medical Image Segmentation NnUnet MedSAM Swin UNet Attention U Net UNETR U Net