Open to full-time & research roles · M.S. CS, Dec 2026
I build agentic AI systems, generative models for biosignals, and the fault-tolerant infrastructure they run on — from diffusion research to production backends.
I'm a Computer Science graduate student and researcher at the University at Buffalo, working where machine learning research meets systems engineering.
My work spans a state-funded multi-agent research platform, peer-reviewed publications at IEEE EMBC, IEEE BSN and the AFRL Quantum AI4 workshop, and a US patent filing — grounded in production experience across backend systems, cloud infrastructure and database internals.
I move comfortably between Python and PyTorch for deep learning and Go or Java for scalable, fault-tolerant backends. I care about models that work outside the notebook, and systems that stay up when things go wrong.
Research Assistant and Teaching Assistant for CSE 474 — Introduction to Machine Learning.
Generative AI, biosignal modelling and fair ML.
A Markovian reasoning engine with conditional diffusion that corrects sensor artifacts, plus a test-time scaling mechanism that adjusts compute on the fly — 91% SOTA accuracy on WESAD/CASE (+14% F1) with 5× fewer labels. The architecture is the subject of a filed US patent.
Personalised EEG-to-image synthesis with adaptive identity evolution via few-shot incremental learning, a unified neural-visual transformer backbone and a continually evolved classifier that prevents catastrophic forgetting — 15–20% accuracy gain over U-Net baselines.
Translating biosignal-driven cognitive-state models into decision-support tooling for operational environments, building on the AtomSigNet reasoning engine.
A chain-of-thought loop that scores synthetic fairness with CLIP and entropy metrics and refactors prompts dynamically — beating SOTA baselines by 4% F1 on WESAD using only 1.4% synthetic data.
A vision-language-kinematic framework for posture coaching that catches joint-level, phase-specific flaws like knee valgus. Video is fused with 3D skeletal dynamics through region-aware anatomical masks; an expanding Mixture-of-Experts routes each motion via Gaussian Mahalanobis selection and learns new exercises class-incrementally without forgetting.
EEG-to-image synthesis with few-shot incremental learning, a unified neural-visual transformer for bidirectional attention, and a temporal masked autoencoder against catastrophic forgetting.
State-funded multi-agent platform for autonomous affective support, with a 500 Hz TimescaleDB pipeline acting as agent memory over live biosignals.
Fault-tolerant distributed key/value store: leader election via randomised timeouts, AppendEntries log replication, disk persistence for crash recovery and conflict-term backtracking.
What I reach for, by layer.
Open to full-time roles, research collaborations and internships. The fastest way to reach me is email.
ritikver@buffalo.edu