I build agentic AI systems that ship at enterprise scale by day, and generative vision models that push research benchmarks by night. Now channeling both into SceneSenseAI, with a PhD next.
BITS Pilani
Indian InstituteI'm Rajarshi: born in Virginia, raised in Bangalore (India's tech capital), and now building AI in New York. That mix of places gave me a simple operating principle: great technology should feel effortless to the person using it, no matter how hard it was to build.
After computer science at BITS Pilani, I finished my M.S. in Computer Science at Stony Brook University, researching generative vision (diffusion models, VAEs, and image restoration) under Prof. Klaus Mueller. That work took me to CVPR 2025, where my team ranked 18th globally in the NTIRE Super-Resolution challenge.
Today I'm a Staff-level AI Engineer at IBM, architecting a patent-pending agentic AI system for the z17 LinuxONE platform and collaborating with IBM Research on agentic workflows for industrial robotics. On nights and weekends, I'm a part-time entrepreneur running SceneSenseAI, building AI teaching assistants that make video course libraries actually searchable.
The through-line: I want to keep doing this at the deepest level possible, which is why the next stop is a PhD.
Computer Vision, NLP, AI, Data Science · Dec 2025
AI, Data Science & Visualization · TA for Object-Oriented Programming · 2024
Generative models, agentic systems, and AI for healthcare & education.
Guitar, singing, movie deep-dives, and building robots that talk back.
From graduate research lab to staff-level engineering at IBM, in under two years.
IBM · LinuxONE / Hybrid Cloud
IBM · LinuxONE Hybrid Cloud Integration & Test
Stony Brook University · VAI-GM Lab, advised by Prof. Klaus Mueller
SceneSenseAI connects to your video library, a Google Drive folder or a YouTube playlist, and turns it into an AI teaching assistant. Learners ask in plain English and get answers cited with exact timestamps; click a citation and the built-in player jumps to that second of that video.
With IBM Research: evaluating 4 agent architectures across 4,000+ real-world asset-inspection scenarios. Targeting ICLR 2027.
Ray R., IBM Z & LinuxONE Community, June 2026. A framework that turns natural-language instructions into fully executed performance tests on IBM Z LinuxONE, orchestrated end-to-end by AI agents.
Ray R., IBM Z & LinuxONE Community, March 2026. Generating realistic synthetic financial transaction data with agentic AI on IBM Granite models: privacy-safe workload testing without touching real customer data.
Ray R., IBM Z & LinuxONE Community, September 2025. Bringing a cloud-native multi-service application to s390x with OpenShift: multi-arch builds, datastore swaps, and architecture-specific fixes.
Chen Z., Ray R., Tomar S.S., Mueller K., et al., Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), NTIRE Workshop, June 2025. Team VAI-GM's MDRCT achieved 30.86 dB PSNR / 0.8942 SSIM, ranking 18th globally.
Generative vision, agentic systems, and the occasional robot: the work fueling my PhD ambitions.
With IBM Research: perception-grounded evaluation of physical AI, benchmarking 4 agent architectures across 4,000+ real-world asset-inspection scenarios.
Led Team VAI-GM in the NTIRE 2025 challenge. Designed the Multi-step Dense Residual Connected Transformer, hitting 30.86 dB PSNR and ranking 18th among global research teams.
Encoded medical images into latent space with VQ-VAE and generated high-fidelity 3D lung CTs via guided diffusion, plus an Orthanc DICOM server for secure storage at scale.
Re-architected the CodeSIM agent loop for 4-bit quantized models (Gemma-2-9B, CodeLlama-7B), keeping HumanEval performance strong under strict hardware and privacy budgets.
View on GitHubFused text, audio, and visual embeddings (concatenation, Hadamard, cross-attention) on the MELD dataset, across unimodal to trimodal setups. The seed that grew into SceneSenseAI.
View on GitHubAt IUDX, IISc: a secure pneumonia-classification pipeline running inside AMD SEV-SNP enclaves with TPM-backed storage and remote inference, ~20% faster than comparable Intel SGX and AWS Nitro stacks.
Whether you're a PhD program, an investor, or a course creator drowning in the same 50 questions: my inbox is open.
CVPR-published, industry-hardened systems experience, and active research toward ICLR 2027. Research statement and references available on request.
Email me about researchSceneSenseAI is live in early access and opening founding pilots with course creators and university instructors. Try the app, then ask for the deck.
Request the deck & a demo