Multimodal AI & Retrieval
Document understanding, text-image retrieval, multilingual embeddings, vector search, OCR-aware reasoning.
Embeddings · CLIP · BERT · Pinecone · cross-document reasoning
Research Scientist @ServiceNow | Master's @MILA AI Institute | Ex-SWE | NIT Warangal
Building multimodal document intelligence, retrieval, evaluation, and AI safety systems for enterprise AI.
Applied Research Scientist at ServiceNow focused on agentic document workflows, multilingual and multimodal search, red-teaming, and scalable ML systems across text, image, audio, and video.
About
Research depth with production instincts.
Applied Research Scientist at ServiceNow building multimodal document intelligence, retrieval, evaluation, and AI safety systems for enterprise AI.
My work spans agentic document workflows, multilingual and multimodal search, red-teaming, and scalable ML services across text, image, audio, and video.
I combine research training from MILA with hands-on product engineering to turn strong ideas into systems that ship.
What this portfolio emphasizes
- Enterprise-grade multimodal retrieval and document understanding.
- Measured impact, not generic skill lists.
- Evaluation, red teaming, and production guardrails for modern AI systems.
- Research outputs that support shipping systems, not notebook-only demos.
Focus Areas
Capabilities that matter for applied AI research and delivery.
Agentic Document Systems
Enterprise document agents that combine extraction, summarization, question answering, and workflow orchestration.
Agentic Frameworks · parsing pipelines · document classification · orchestration
Evaluation, Red Teaming & AI Safety
Adversarial testing, prompt-injection defense, safety datasets, and robustness evaluation for multimodal agents.
Red teaming · prompt hardening · sanitization · benchmark design
Production ML Platforms
Scalable pipelines that connect research models to services, monitoring, and real user workflows.
Kubernetes · Docker · GraphQL · React · model serving · microservices
Experience
Recent work across applied research, enterprise AI, and product systems.
Applied Research Scientist
- Built a document intelligence agent within ServiceNow's agentic ecosystem for summarization, Q&A, information extraction, classification, and cross-document reasoning across enterprise documents.
- Implemented multilingual and multimodal vector search across 10+ languages and 20+ document types, deploying Kubernetes microservices for parsing, captioning, and content embedding.
- Collaborated with NVIDIA on dataset curation, fine-tuning, and benchmarking across OCR, retrieval, and Q&A model variants.
- Extended multimodal services into audio and video workflows with efficient embedding generation and content understanding pipelines.
- Engineered a semi-supervised labeling pipeline using HDBSCAN clustering, reducing manual annotation by 60%.
- Led fraud document detection research with a five-agent modular AI system across layout, semantic, anomaly, tampering, and metadata signals, reaching 92% accuracy on synthetic fraud datasets.
- Contributed to red-teaming and adversarial evaluation of multimodal document agents, including prompt-injection and hidden-instruction attacks, and added guardrails through prompt hardening and sanitization.
Student Researcher
- Built RoboPDF AI for legal query processing by fine-tuning LLaMA 2 on a Canadian legal corpus with QLoRA and a custom RAG pipeline, improving retrieval accuracy by 40%.
- Co-authored research on LLM evaluation for mediation and AI-based dispute resolution and published the
text-moderatenpm package for LLM content moderation.
Software Engineer II
- Designed DocChat, an enterprise document retrieval system using BERT embeddings and Pinecone vector search for natural-language query processing.
- Built an unsupervised recommendation engine using text embeddings, reducing average ticket resolution time by 23%, and shipped supporting product features with React, Java, and GraphQL.
Selected Work
Three projects that best represent my current research-plus-product profile.
Flagship project
LLM-Aug-MoE
Built a multilingual mixture-of-experts architecture to improve low-resource language performance without scaling a single monolithic model.
Combined a compact base model with task-specific experts through cross-attention and selective layer freezing to keep training efficient.
Improved low-resource multilingual performance by 28% while reducing compute cost by roughly 60%.
Multimodal systems
Vision Augmented Large Language Models (VA-LLM)
Designed a multimodal reasoning system for joint text-image understanding instead of a pure text-only LLM workflow.
Integrated CLIP ViT with StableLM via cross-attention and packaged the workflow with Docker and Hugging Face Hub for reproducible experimentation.
Turned a research prototype into a repeatable multimodal experimentation stack rather than a notebook-only demo.
Live deployment
NHL Expected Goals (xG) Prediction & Live Deployment
Built an end-to-end expected-goals pipeline for hockey analytics with real-time data and live model serving.
Shipped a Dockerized Flask API, Streamlit dashboard, NHL API ingestion layer, and CometML-backed model registry for experiment hot-swapping.
Demonstrated production-style model serving, monitoring, and real-time prediction loops in a research environment.
Publications
Research output anchored in evaluation, retrieval, and real-world deployment questions.
Bridging Modality Gap: Enhancing Document Retrieval with Multimodal Embeddings
Master's thesis exploring multimodal retrieval for document intelligence.
Let's Evaluate Step-by-Step: A Robust Evaluation Method of LLM Applications
Framework for more reliable evaluation of applied LLM systems.
Robots in the Middle: Evaluating LLMs in Dispute Resolution
Study of LLM behavior in mediation and AI-based dispute resolution settings.
Leveraging AI for Natural Disaster Management: Takeaways from the Moroccan Earthquake
Applied AI analysis for humanitarian and crisis-response contexts.
Efficient Detection of Disguised Faces from Low-Quality Surveillance Footage
Computer vision work focused on robust detection under low-quality surveillance conditions.
Education
Built on strong ML training and a computer science foundation.
Graduated Dec 2024
MILA / Universite de Montreal
Master's in Computer Science, Machine Learning specialization
Grade: 4.2 / 4.3 · Affiliated with UdeM / McGill
Graduated May 2021
National Institute of Technology, Warangal
B.Tech in Computer Science & Engineering
Strong foundation in algorithms, systems, ML, and software engineering.
Contact
Open to research, multimodal AI, and product conversations.
The fastest way to reach me is by email or LinkedIn. If you want the latest version of my background, use the resume link below.