A world-modeling agent for the ARC Prize 2026 ARC-AGI-3 competition, built to explore unseen interactive grid games, infer rules and goals, and plan actions. Result: measured on the public offline engine, Curio made progress on 13 of 25 games, fully solved ft09 across all 6 levels below the human action baseline, scored 8.74 on the all-25 efficiency scorecard, and reached 0.087 held-out generalization.
Open projectA Kaggle Orbit Wars competition agent and self-play tournament harness, with engine-exact planet and fleet simulation, forward-search opening decisions, public-opponent benchmarks, and reproducible ranking tools. Result: in a 1,650-game local tournament against public 1000+ leaderboard agents, Hybrid reached an 86.7% overall win rate and beat the top public agent 57% head-to-head.
Open projectA pure-PyTorch world-model research lab where a continuous rocket lander learns a latent simulator, trains its policy inside imagined rollouts, persists checkpoints, and shows real versus dreamed behavior in a local dashboard. Result: the committed Apple MPS smoke run produced an eval return of -118.08 after 80 steps, dream-to-real error of 0.137, and a 46,287-parameter model.
Open projectA program search system that improves code by generating candidates, testing correctness, benchmarking performance, and saving reproducible evidence for each run. Result: on the packaged Levenshtein benchmark, an 8-attempt verified run improved mean runtime from 0.001421s to 0.000274s, a 5.18x speedup, with the winner rerun passing verification.
Open projectA pure-PyTorch research project for plastic-memory sequence models: local attention, bounded differentiable memory, surprise-gated writes, and ablations that test whether models can learn what to remember and forget. Result: the local Apple Silicon sanity run reached 0.6125 accuracy on the changing_facts overwrite task over 160 held-out examples after a 100-step checkpoint.
Open projectA PyTorch research repo for predicting how a cell population changes after an unseen gene perturbation, with a controllable synthetic benchmark, held-out combinatorial tests, figures, and a short technical note. Result: on the held-out synthetic split, the model reached MSE 0.0386 and mean gene correlation 0.871, beating direct-shift MSE 0.1090 and mean-shift MSE 0.1428 baselines.
Open projectA research project that tests whether GPT OSS models can improve their answers by generating several options, checking them with verifiers, and saving the results as clear reproducible evidence.
Open projectA custom reinforcement-learning environment and PPO agent for landing a 2D rocket under thrust, torque, fuel, and strict landing constraints, with expert imitation, curriculum learning, and checkpointed normalization.
Open projectA PyTorch computer-vision project that classifies Galaxy Zoo 2 images into elliptical, spiral, and edge-on disk galaxies using ResNet18 and a VRAM-safe high-utilization training loop.
Open projectA Django agent application for browser-based workflow automation, combining LLM-driven task planning, Browser Use automation, live monitoring, WebSockets, user accounts, and an admin dashboard.
Open projectAn AI-powered Django website builder that turns chat-driven design iterations into responsive HTML, CSS, JavaScript, and Bootstrap sites that can be published to named site URLs.
Open projectA private Django subscription web app with Free and Plus tiers, usage tracking, search history, activity analytics, responsive account dashboards, and integrated payment-processing foundations.
Private repo