mmbert

Local-first privacy guard: anonymize your documents before sharing with LLMs.

🌳 Groot 1 — open typed-decision model. Calibrated choice/score/yes-no answers over 64k-token states, System One API compatible, with a pure Bend engine. Beats Jev on zero-shot Folha news (87.8% vs 85.8%) and typed decisions (75.3% vs 73.7%).

Typed decisions from one encoder pass. Open weights, Apache 2.0.

Anonymize Italian legal text locally with reversible PII replacement to maintain data privacy while using AI models.

Free Laya AI Model API Gateway — sub-35ms deterministic inference, prompt guard, and classification API

Stacked multilingual guardrail (mmBERT + Cohere c4ai LLM judge) for crisis detection in youth chat. 1st of 77 teams at the Bell x Mila x BuzzHPC x Kids Help Phone AI Safety Hackathon. F1 0.906, recall 0.969.

Локальный движок типизированных решений (choice/score/noul) на Go + ONNX Runtime: ~322M параметров, один forward pass на батч вопросов, GPU/CPU, без Python в рантайме. Плагин и навык ZCode: измеренная экономия −48% входных токенов и −58% пика контекста на большом логе.

Decidra - 决策原子化的多语言通用决策平台：把判断拆成类型化问题，一次前向传播得到带概率的答案，用置信门控决定自动执行还是升级人工

Multilingual personal-information detector: names, contacts and ID numbers as spans, plus document sensitivity (fine-tuned mmBERT-base)

Convert and deploy laya — the open, Jev-compatible System One decision model — on Rockchip RK3588 and NVIDIA Jetson Orin. Conversion recipes, measured latency, and a reproducible acceptance protocol. RK3588 509 ms · Jetson Orin NX 18.9 ms · 16/16.