Niansia

An in-browser playground for adversarial attacks: draw or pick an MNIST digit, attack a CNN with FGSM, PGD or targeted PGD under an L∞ budget, watch the prediction flip step by step, and compare it with a PGD adversarially trained model through gradients, transfer and decision maps of input space.

Result: On 2,000 MNIST test images under PGD-40: the standard model falls from 98.95% clean accuracy to 0.2% at ε = 0.2, while the adversarially trained model keeps 92.2% at ε = 0.2 and 85.3% at ε = 0.3 (97.92% clean).

Low-light image enhancement that predicts a luminance-guided bilateral grid of Zero-DCE curves and colour matrices from a thumbnail of the whole image, slices it at full resolution, and cleans noise and detail with a lightweight NAFNet refiner.

Result: On 20 held-out NTIRE 2025 pairs with the official scoring code: 24.57 dB PSNR / 0.840 SSIM (24.63 dB with test-time augmentation), against 16.48 dB for the original Zero-DCE course pipeline and 20.91 dB for Zero-DCE trained on the same data; component ablations, released weights, 2.1 M parameters, 24-megapixel images on an 8 GB laptop GPU.

An Agent Skill for creating original Taiwan GSAT practice exams.

Result: Subject-specific guidance, source and answer checks, layout validation, exam templates, and public installation instructions.

A simulation-first research platform for deterministic, reproducible Windows driver reliability experiments.

Result: Canonical Case IR, seeded scheduling, an append-only SQLite journal, exact signatures, 3/3 simulated replay, recorded G3, E1, E2, and minimization artifacts, semantic verification, and multi-platform CI.

An evidence-first Agent Skill for adversarial scholarly novelty audits.

Result: Claim maps, provider-normalized records, evidence-bound prior sets, temporal states, citation-graph bridges, deterministic report invariants, three public case-study classes, and exploratory measurements from 82 reviewer-annotated cases.

A vendor-neutral release-evidence layer for deployable AI models.

Result: Reproducible release records, explicit promotion policy, provenance, regression localization, and an independent verification path.