A visual guide to an AI project that builds specialist neural networks from scratch — and makes them small enough to run almost anywhere.
The core is a tiny neural-network toolkit. Around it: a deployable fraud service, a product site, and a growing collection of research experiments.
Normalize numeric inputs. Encode recurring signals such as time of day.
Calculate gradients directly, then update weights with SGD or Adam.
Use validation loss, early stopping, and task-specific benchmarks.
Run locally, serve through the API, or export selected experiments to C or ONNX.
This is the saved KestrelNet model, running in JavaScript. Adjust a transaction and see all three class probabilities update.
Other inputs stay fixed: 2pm on day index 2, three transactions in 24 hours, a one-year-old account, and no previous fraud history.
Preparing local inference.
Educational demo. The fraud model was trained on synthetic data; its probabilities are model outputs, not calibrated real-world risk estimates.
Five published GnaniNet notebooks. Results retrieved from Kaggle execution logs on October 7, 2026.
Download the verified resultsStrong classification results in compact CPU models. These examples compare our measured results with baselines listed in our notebooks.
99.06% accuracy with 163,788 parameters.
Above the listed logistic regression and random forest baselines at 97%. Their parameter counts were not reported.
94.88% accuracy with 15,416 parameters.
About 13× fewer parameters than the listed LSTM (~200K), with higher accuracy (93.9% for that baseline).
94.16% accuracy with 1,576 parameters.
About 571× fewer parameters than the listed LeViT (~900K); LeViT has higher raw accuracy at 99%.
Baseline figures and approximate model sizes are authored notebook comparisons, not baselines rerun by Kaggle. These records establish our model outputs, not a global #1 ranking for accuracy per parameter. Parameter count does not measure FLOPs or lowest computation.
The separate fraud demo above uses the local 1,059-parameter model. Its repository-reported 91.6% accuracy is not one of these five Kaggle results.
Separate competition tracks explore how compact models and game strategies tackle different kinds of reasoning.
Build small ONNX networks for ARC-style grid puzzles. The best recorded Kaggle submission scored 806.30 on both public and private results (April 27, 2026). Its submission description records 53 tasks.
competitions/neurogolf/View NeuroGolf on KaggleA substantial game-agent research track with strategy variants, self-play training, opponent analysis, and a local viewer for game replays.
competitions/orbit-wars/Completed submission on April 15, 2026: public score 0.90820, private score 0.90900. The submission description records a 14,851-parameter FCNet. Competition scores are separate from the notebook accuracies above; neither this entry nor NeuroGolf ranked first on the final leaderboards.
A research toolkit growing into a product.
The repository brings together model code, saved weights, benchmarks, an authenticated FastAPI service, and a Next.js website. This explorer is a standalone guide to those pieces.