Simba Shi
I study CS and Math at Yale, and I love building AI and robots.
Previously, I cofounded ReefSound, an AI and robotics startup for ocean monitoring scaled across 7 countries, featured by The Independent, NASA, United Nations, and National Geographic.
My current research is in machine learning and robotics, spanning multimodal LLMs, RL, mechanistic interpretability, generalist robot policies, and I'm interested broadly in intelligence in both its digital and biological forms.
In another life, I'm a creative writer. Here, you'll find small, scattered fragments of my life & work.
01 — cool places I've worked
Experience
Tractable
May 2026 – PresentMachine Learning Intern · $1B valuation, Series E
Scaling ML edge inference for hardware at one of UK's top computer vision unicorns.
Yale Graph and Geometric Learning Lab
Mar 2026 – PresentStudent Researcher
Researched post-training and RL for multimodal time-series LLMs for quantitative finance. Mentored by Rex Ying.
SPAR Research
Feb 2026 – PresentAI Alignment Research Fellow
Activation steering and chain-of-thought monitorability with MIT and ex-Meta researchers.
ReefSound
Jul 2023 – Aug 2025Co-Founder
Audio AI & underwater robotics for automating ocean monitoring. Deployed in 7 countries. Featured by The Independent, NASA, National Geographic.
Greyparrot
Jul 2024 – Sep 2024ML Engineer Intern · Series B, $30M raised
Computer vision and YOLOv8 inference pipelines for industrial recycling.
CVSSP Lab, University of Surrey
Jun 2023 – Feb 2025Research Intern
GANs for zero-shot image translation; seq2seq transformers and graph convolutional networks for audio captioning.
02 — some things I built
Research
Prediction Is Not Physics: Learning Conserved Quantities in Neural Simulators
Trained neural nets to discover Hamiltonian physics; showed that low rollout MSE and physical conservation are orthogonal: diffusion energy drift is 7,500–36,000× ground truth.
Optimized Quantum Cat Qubit Reed-Solomon & Tornado Codes
Novel quantum error correction algorithm; Tornado architecture combining Reed-Solomon codes and repetition stabilizers for cat qubits.
Safe Pedestrian Navigation App using NAMOA*
Reduced NAMOA* pathfinding runtime from O(N³) to O(N log N) via skyline pruning and cardinality bounding.
Towards a Novel Multi-Loss CycleGAN for Full-to-Flat Colour Transformation
GAN pipeline for zero-shot full-to-flat colour image transformation; CycleGAN augmented with depth geometry (InceptionV3), sketch extractor, and CLIP semantics losses.
Mycoflo: Pleurotus tuber-regium in Heavy Metal Filtration and Machine Learning for Water Potability Testing
Mycofiltration using native Nigerian fungi + XGBoost/Random Forest ensemble for water potability; 63.75% copper reduction, 94% classification accuracy. Featured by United Nations; The Earth Prize 2024 Finalist (Top 10 of 1000+).
03 — fragmented memories