About

  • Schulich Leader Scholar at McMaster University, one of Canada's top 50 entrepreneurial engineers that was awarded $120,000.
  • Wrote the first native Blackwell (tcgen05) training backward for the gated-linear-recurrence family (Mamba-2, GLA, KDA, Gated DeltaNet).
  • Designed a novel training objective for latent world models that nearly doubles valid prediction time with one loss term and zero added parameters.
  • Trained the supervised and reasoning layers of a labor-prediction model at E3A Healthcare that outperforms the clinical SOTA by 3x.
  • Built a novel linear-attention sequence model from scratch in C++17, trained on clinical data to predict hypoglycemia 60 minutes ahead, and proved its kernel runs bit-identically on a $4 chip.
  • Engineered a biologically inspired recurrent vision model that outperformed MIT's CORNet-S in adversarial robustness across MNIST, CIFAR, and ImageNet100.
  • Worked with Synaptive Medical and surgeons to develop a surgical-vision system trained on live operating-room data, achieving 95% tool segmentation accuracy for real-time surgical guidance.
Lethe
2026

Lethe

Kernel Engineering Machine Learning CUDA

First hand-written native Blackwell (tcgen05) training backward for the gated-linear-recurrence family (Mamba-2, GLA, KDA, Gated DeltaNet), verified against an fp64 oracle and bit-for-bit deterministic. Includes a 12-gate contract verifier that audited 2,638 LLM-generated kernels a public system had accepted and falsified 62.1% of them, plus the first open Triton Mamba-3 MIMO backward.

RDR
2026

RDR

World Models Machine Learning Model-Based Control

A one-term training objective for latent world models that free-runs the model during training, decodes every rollout latent, and penalizes it against ground truth. Adds zero parameters and changes no architecture, so setting its weight to zero recovers the standard objective exactly and every result is a one-flag A/B. Nearly doubles valid prediction time on chaotic Kuramoto–Sivashinsky at an identical 193,568 parameters, 10 of 10 preregistered configurations, with the margin growing as the latent widens.

CGM-on-Chip
2026

CGM-on-Chip

Machine Learning Embedded Systems Healthcare

Linear-attention architecture (S4D + OSDN) reimplemented from scratch in ~5,500 lines of pure C++17: autograd engine, optimizer, multi-threaded trainer, embedded FP32 kernel, and a triple-redundant correctness proof. Trained on real clinical data; runs on a $4 ESP32-S3 in 20% of its RAM to predict hypoglycemia 60 minutes ahead.

From Pixels to Precision
2023

From Pixels to Precision

Surgical Vision Healthcare Deep Learning

Partnered with Synaptive Medical and Robarts Research Institute to build surgical-vision AI system achieving 95% tool-segmentation accuracy for real-time guidance. Bronze Medal at Canada-Wide Science Fair for fluorescence-guided surgery support with <0.12s per image processing.

Adversarial Robustness
2025

Adversarial Robustness Research

ML Research Adversarial Robustness Computer Vision

Engineered a bio-inspired recurrent vision model achieving 97.82% adversarial robustness, outperforming MIT's CORNet-S across MNIST, CIFAR, and ImageNet100. Integrated learnable prefiltering, gated recurrence, and denoise-scaling without requiring any adversarial training.

Sentinel
2025

Sentinel

Developer Tools Embedded Systems Language Server

First hardware-aware static analyzer for embedded systems performing RAM/Flash estimation, pin-mapping, and 50+ validations before compilation. Built a custom Language Server Protocol integration with VS Code for real-time hardware-intelligent error detection. Won Best Developer Tool at Hack Western 12; 1K+ installs across the VS Code Marketplace and Open VSX.

AdaptVR
2026

AdaptVR

VR Rehab Healthcare Embedded VR

Meta Quest 2 native Unity app for progressive muscle disease rehab. World Labs Gaussian-splat environment loaded as the world, hand-tracked butterfly catching via OVRSkeleton wrist bones, configurable spawn count and motion zone, overextension guardrail, and a singleton-driven score counter + session timer.

Adaptive Learning Robot
2025

Adaptive Learning Robot

Robotics LLMs Embedded Systems

Built a dual-system robot using Cohere LLMs and real-time search that learns new skills instantly from natural language without pre-training. Placed Top 32 out of 256 projects at Hack the North with 2 custom 3-DOF robotic arms executing smooth motions from messy instructions.

Zephyr
2024

Zephyr AI Assistant

Personal AI Productivity Multi-LLM

Keyboard-activated AI assistant with dual interface combining instant popup access and powerful dashboard for seamless workflow integration. Leverages Groq's Llama 3.1 8B for voice commands, intelligent memory, and controls Spotify, weather, news, calendar, and project tracking.

DermAI
2024

DermAI Medical Assistant

Medical AI Healthcare RAG

Medical AI platform combining DenseNet121 classification trained on 19,000+ DermNet images with RAG-powered chatbot for comprehensive triage. Provides instant dermatological assessment and personalized health guidance through dual AI architecture.

SurgicalToolAR
2024

SurgicalToolAR

Mixed Reality Healthcare Surgical Vision

Augmented reality application for surgical tool visualization and training. Interactive AR experience for medical education and surgical workflow enhancement.

Gesture Controlled Robotic Arm
2024

Gesture Controlled Robotic Arm

Computer Vision Robotics Embedded Systems

Computer vision-controlled robotic arm responding to hand gestures in real-time using MediaPipe and Arduino. Demonstrates natural human-robot interaction with 6-DOF movement mapped from gesture recognition.

2026 · Preprint

Rollout-Decoded Reconstruction for Latent World Models

Rishi Shah, Rishav Shrestha · DOI 10.5281/zenodo.21894547

A latent world model's decoder is trained on states anchored to observations and deployed on the model's own free-running rollout. RDR is one loss term that closes the gap: free-run during training, decode every rollout latent, penalize the error against ground truth. It adds no parameters, so every comparison is a one-flag A/B. On chaotic Kuramoto–Sivashinsky it raises valid prediction time from 3.87 to 6.97 time units at an identical 193,568 parameters, winning 10 of 10 preregistered configurations.

2026 · Preprint

A Contract-Grade Verifier for LLM-Generated GPU Kernels, and a Native Blackwell Backward for the Gated-Linear-Recurrence Family

Rishi Shah, Rishav Shrestha · DOI 10.5281/zenodo.21563213

Twelve adversarial gates that check GPU-kernel correctness properly, several of them tolerance-free. Aimed outward, the verifier audited 2,638 LLM-generated kernels a public system had already accepted and found 62.1% carrying a contract violation; aimed inward, it certified the first native Blackwell (tcgen05) training backward for the gated-linear-recurrence family.

May 2026 – Aug 2026Shenzhen, China

Machine Learning Engineer

E3A Healthcare

Led the supervised layer and set the architecture direction of a longitudinal model that reads at-home fetal monitoring and returns a full posterior over days until delivery. The model predicts term delivery to within 3 days against 11 days for the clinical SOTA, and halves the obstetric calendar's preterm MAE.

  • Authored the first design of the cohort mixture-of-experts and its learned gate, and originated the sub-agent architecture now on the model's roadmap.
  • Built the reasoning layer and the coupled-likelihood objective behind the preterm win, 2.02 days better with the same sign on all ten seeds.
  • Wrote the evaluation methodology the project now runs on, and put GPU compute in every engineer's hands by automating multi-region GPU cluster provisioning on GCP.
Oct 2025 – Apr 2026Hamilton, ON

Research Engineer

McMaster Automotive Resource Centre (MARC)

Cut battery voltage-model error 84% (12.82 mV to 2.03 mV RMSE) with a physics + LSTM hybrid, then carried it through ONNX to MATLAB and Simulink with machine-precision parity. Co-authoring a paper on agentic-AI-accelerated state-of-charge estimation built on a from-scratch EKF and SVSF.

Apr 2024 – Aug 2025London, ON

Machine Learning Researcher

Western University, Vision & Robotics Lab

Designed a biologically inspired recurrent vision model that reached 97.82% adversarial accuracy on MNIST without a single adversarial training example, outperforming MIT's CORNet-S by 12 points across MNIST, CIFAR-100, and ImageNet100.

Nov 2022 – May 2024London, ON

Student Researcher, Medical Imaging

Robarts Research Institute

Partnered with Synaptive Medical on surgical-vision AI: a U-Net at 95% DICE and sub-0.12 s per frame on live operating-room footage, plus a marker-based AR guidance prototype.