About
I am a PhD student and AI Research Engineer based in the UAE, with three years of experience developing and deploying computer-vision, multimodal, and real-time inference systems. I have built vision pipelines operating at 76 FPS, sub-second streaming ASR systems, and scalable model-serving infrastructure using PyTorch, C++, Docker, and cloud platforms.
My research centres on scientific foundation models and world models, autonomous agents for personalized medicine, AI for AI (alignment, safety, and mechanistic interpretability), and the scalability stack that lets these systems run at the edge of care.
Alongside research, I care about the engineering that makes it real: inference optimization, knowledge distillation, agent-native memory, and hardware-aware deployment inside clinical and discovery workflows.
I welcome opportunities in academia and research labs — including full-time roles, visiting fellowships, and internships. A copy of my CV is available here.
You can reach me at abhijit.das@mbzuai.ac.ae.
Education
Mohamed bin Zayed University of Artificial Intelligence (MBZUAI)
Ph.D. in Machine Learning
Maulana Abul Kalam Azad University of Technology
B.Tech in Computer Science and Engineering
- Thesis: Attention Capsules for Robust Medical Imaging (100/100).
- Adding attention inside vanilla CapsNet routing yields faster routing and robust performance at scale.
- GPA: 9.2 / 10.
Experience
MBZUAI
Research Engineer I
- — Faithful reasoning and hallucination control in medical vision-language and foundation models via counterfactual, anatomy-guided, and contrastive decoding (PCCD, CAST, Transductive Conformal Decoding).
- — Risk-controlled, uncertainty-aware adaptation for trustworthy clinical deployment: atlas-aware and open-vocabulary conformal prediction with online OOD detection (CARTA, OpenCP, PROTON).
- — Annotation- and data-efficient learning under distribution shift for medical imaging: structured active learning under correlated outputs and black-box foundation-model prompting.
- — Mechanistic interpretability of VLMs and self-improving agentic systems for automated evaluation and tool use (State Blindness, ViTA, GEM).
MedOS Limited
Co-Founder & CTO
- — Founding technical lead setting the AI research and engineering direction for a clinical AI platform.
- — Building the first dedicated clinical AI assistant, with FDA approval as a priority track.
Innovxcare AI
AI Engineer
- — Developed Spark App for real-time radiology reporting in an expert-in-the-loop framework, aligning textual dictations with imaging data.
- — Built real-time ASR pipelines on open-source foundation models with sub-second streaming latency and domain adaptation.
- — Optimized and packaged inference pipelines using C++, WASM, and WebGPU backends for resource-constrained clinical environments.
Wipro GE Healthcare
AI Researcher
- — Designed a Quantum-inspired Convolutional Neural Network (QCNN) with improved robustness to correlated noise for CT/MR denoising.
- — LLM finetuning for MR/CT Radiology Information System series-description understanding and reasoning.
- — Symbolic regression to detect and localize internal organ spans from external landmarks in multi-view RGB streams.
OnFinanceAI
AI Engineer — Founding Team
- — Built an in-house VLM for financial chart understanding and reasoning based on the GLM vision model.
- — Team lead of Voice-2-Compliance, India's first generative AI-powered voice-based compliance solution for BFSI, adopted by the National Stock Exchange.
Jio Institute
Research Assistant
- — Label-efficient neural networks.
- — Complementary geometric representations (spectral and Euclidean spaces) in CNNs.
Feinberg School of Medicine, Northwestern University
Research Fellow
- — AI for liver health (HCC, cirrhosis, interventional therapy), supervised by Dr. Ulas Bagci.
- — Developed a real-time endoscopy polyp-detection system achieving 76 FPS.
Skills
- Efficient ML
- Real-time inference, model compilation, latency and throughput optimization, resource-constrained deployment.
- Research
- Foundation-model pretraining, simulative world models, conformal prediction and uncertainty/risk control, test-time and OOD adaptation, faithful multimodal reasoning and hallucination mitigation, mechanistic interpretability.
- ML Frameworks
- PyTorch, TensorFlow, Hugging Face, vLLM.
- Systems
- Python, C++, Docker, FastAPI, AWS, GCP, WASM, WebGPU.
- Model Optimization
- Quantization, pruning, knowledge distillation, mixed-precision inference, ONNX / TensorRT / TFLite.
Awards
Featured at Digestive Disease Week 2024
Recognized for contribution to real-time sessile serrated polyp detection on video using DAPO-Det.
Contact
- Phone
- +91 8101969424
- Based in
- MBZUAI, UAE
- Elsewhere
- Google ScholarGitHubLinkedInCV (PDF)