An isometric browser city-builder with a deterministic simulation engine.

A clean-architecture rebuild of isometric-city: zone land, build roads and rail, and watch a full economy grow with autonomous cars, trains, boats, planes, a day/night cycle, and disasters. A pure TypeScript engine with a seeded RNG drives a five-layer canvas renderer.

A spatial interface for AI conversations on an infinite canvas.

Moves beyond linear chat with a 2D workspace where text, image, post-it, and file nodes can be freely arranged. Supports live multiplayer cursors, AI chat and image generation, version history, and shareable canvases.

A feedforward neural network built from first principles in CUDA.

A 3-layer classifier written entirely in CUDA C, with custom GPU kernels for forward and backward passes, an Adam optimizer running on the GPU, and hand-rolled matrix operations accelerated with cuBLAS and cuDNN. No deep learning framework involved.

A full-stack drug discovery platform spanning gene to molecule.

Five integrated modules covering codon optimization, protein structure prediction, molecular docking, protein interaction networks, and external database search. Built with a React frontend and a Node/Python backend with background job queuing, team collaboration, and AI-generated research reports.

Predicts how genetic perturbations and drug treatments reshape a cell.

Takes a natural-language query like "What happens if I knock out TP53 in IFNy cells?" and predicts RNA expression across 19,000+ genes and 500+ proteins, then runs pathway enrichment analysis and generates AI hypotheses with PubMed citations. Includes interactive 3D cell and volcano plot visualizations.

RL fundamentals built from scratch through Tic-Tac-Toe.

A progressive study of reinforcement learning: starting with value iteration over the full game state space, then Monte Carlo learning and Q-learning with self-play, then a deep Q-network with experience replay. Each method reveals a different trade-off between sample efficiency, stability, and scalability.

A macOS menu bar app that enforces the 20-20-20 rule for eye strain.

Every 20 minutes, a fullscreen overlay prompts a 20-second break. Built with SwiftUI and a native AppKit backend, with a timer state machine, multi-display support, system notifications, and keyboard shortcuts.

De novo enzyme design for degrading environmental toxins.

An end-to-end pipeline for designing novel enzymes targeting PFAS, microplastics, and biofilms. Analyzes a target substrate's chemistry to generate structural constraints, runs a protein backbone diffusion model to propose candidates, designs sequences using a ligand-aware neural network, then validates folding with AlphaFold2. All stages are orchestrated on an HPC cluster.

A dual-stream neural network for predicting molecular properties.

Combines a convolutional network reading molecule string representations with a graph neural network reading the underlying molecular graph, then fuses them to predict multiple properties simultaneously across five benchmark datasets covering solubility, lipophilicity, blood-brain barrier penetration, and toxicity.

Reconstructs evolutionary relationships from DNA sequences, from scratch.

Uses Smith-Waterman local alignment to compute pairwise genetic distances between sequences, then builds an evolutionary tree by iteratively merging the closest species using the UPGMA algorithm. Includes an interactive visualization with rectangular and radial tree layouts.

A research-grounded protein design agent built on Claude and ESMFold.

Combines an AI agent with a curated library of 50 protein design papers and a structure prediction model. Given a design prompt, it retrieves relevant research to ground its reasoning, generates candidate sequences, folds them, and visualizes the result in 3D. Available as both a CLI and a web API.

A ground-up PyTorch reimplementation of AlphaFold 3.

Follows Abramson et al. (Nature 2024) end to end, cross-referencing Boltz and OpenFold3 for paper-faithful detail. Covers the unified token scheme that treats residues and ligand atoms alike, the 48-block Pairformer trunk, and the key shift from AF2: replacing the structure module with a diffusion process that denoises directly on atomic coordinates, enabling a single model to handle proteins, RNA, DNA, and small molecules.

A pure-PyTorch SE(3)-Transformer built from scratch.

A module-by-module reimplementation of Fuchs et al. (NeurIPS 2020). The core idea: a network that is exactly equivariant to 3D rotations and translations by design, not approximated through data augmentation.

A clean, well-documented PyTorch implementation of AlphaFold 2.

Follows Jumper et al. (Nature 2021) end to end: the Evoformer trunk that jointly processes multiple sequence alignments and pairwise residue relationships via triangle attention and multiplicative updates, then a structure module using Invariant Point Attention to iteratively build a 3D protein backbone from predicted torsion angles and frames.

The original Transformer architecture, implemented from scratch.

A faithful implementation of "Attention Is All You Need" (Vaswani et al., 2017) for machine translation, including scaled dot-product attention, multi-head attention, sinusoidal positional encodings, and the full encoder-decoder stack. Reproduces the paper's training setup including the warmup learning rate schedule and label smoothing.

RNN architectures traced through the history of their development.

Implements vanilla RNN, LSTM, peephole LSTM, coupled-gate LSTM, and GRU in the order they were introduced, training each on character-level language modeling. Each variant isolates a specific architectural decision so you can see exactly what each gating change fixes.