Engineering logs, benchmark results, research findings, and honest progress reports. Written from git commits, not a content calendar.
Major site expansion — enterprise landing page, compliance matrices (EU AI Act, NIST, OWASP), full-page Polly chat with slash commands, test suite dashboard, tools catalog, workflow gallery, and contact forms. Unified design system across all pages. Sitemap went from 35 to 84 URLs.
Read more →Training data audit revealed 84% of records labeled "unknown" and 134 labels with fewer than 10 samples. Consolidated 254 raw labels into 24 families. Accuracy on the placeholder model dropped from 66% to 43% — because the task is now real instead of memorizing noise. Qwen fine-tune on Kaggle next.
Read more →Pricing page shipped ($49/$199/$499), benchmark suite with 91-attack corpus, spectral sweep across 8 configs, and the 233K multi-view SFT dataset pushed to HuggingFace. Research page updated with March 31 verified numbers.
Read more →How the 14-layer pipeline provides runtime oversight for hub-and-spoke multi-agent architectures. Three pillars of defense mapped to specific pipeline layers: input validation, planning validation, and behavioral analysis.
Read more →