Omkar Thawakar

PhD Researcher | Multimodal AI | Video Understanding | LLMs & Agents

PhD researcher at MBZUAI, working on multimodal reasoning, video understanding, large multimodal models (LMMs), and self-evolving AI systems, with strong focus on real-world deployment.

Highlights

[ECCV 2026] 1 paper accepted.
[CVPR 2026] 3 papers accepted.
[ICLR-SLLM 2025] Spotlight (Top-2%) for MobiLLaMA.
[CVPR 2025] Highlight for All Languages Matter (LMM Evaluation).
[Impact] 300K+ HuggingFace downloads across models.
[Award] Khalifa Fund Entrepreneurship Competition Winner (250K AED).
[Award] Sandook Al Watan Student Project Competition Winner.

Spotlight Research

MobiLLaMA (ICLR 2025)

Accurate & Lightweight Fully Transparent GPT. 200K+ Downloads.

Read Paper

LlamaV-o1 (ACL 2025)

Rethinking Step-by-Step Visual Reasoning in LLMs.

Read Paper

Recent Projects

Nutrigenics.Care

Personalized AI nutrition and lifestyle platform for health monitoring

Nutrigenics.Care is an intelligent clinical nutrition platform that builds personalized dietary recommendations. Powered by our proprietary Nutrition-GPT model, it offers patients and wellness enthusiasts real-time diet analysis, macro-nutrient optimization, and lifestyle advice cross-referenced with medical research.

Nutrition-GPT Clinical Validation Diet Coaching Privacy-First

Integrates with clinical guidelines for medically sound nutritional feedback.
Generates personalized daily meal plans and targets based on biometric data.
Uses interactive AI chat to answer questions about ingredients, recipes, and allergies.
Maintains secure health records on-device or within private cloud sandboxes.

Grants & Funding

Nutrigenics.Care has secured over $100K+ in grant funding and is an active member of the Microsoft Founders Hub, receiving $150K in technical and infrastructure support.

VisQ (Visual Query)

iOS application for composed image and video retrieval on iPhone

VisQ brings reason-aware visual retrieval to iPhone with an on-device Qwen3-VL-2B Core ML runtime. Users can search personal media with natural language or run composed retrieval using a reference image + edit prompt, then inspect "Why This Matched" explanations powered by the model's reasoning capability.

On-device AI Composed Retrieval Explainable Results Privacy-First Offline-First

Indexes local photos and videos directly from the iPhone photo library.
Supports text search and reference-image-guided retrieval with scene edits.
Surfaces human-readable match reasons and visual explanation chips.
Keeps embeddings, ranking, and inference on-device for privacy-preserving search.

App Store Download Now GitHub Source code

Built from research

VisQ is based on our recent research work CoVR-R: Reason-Aware Composed Video Retrieval, translating reason-aware composed retrieval into a practical iPhone app for local-first multimodal search.

Available now on the Apple App Store.

arXiv Research GitHub