We build the local-first AI runtime behind multiple products.
D4DA is the company and Google Cloud account domain. Its core software is a browser-native AI training, evaluation, and inference stack for privacy-sensitive workflows. Valera is the legal review application built on that stack.
Core infrastructure
Local-first training, evaluation, distillation, benchmark, inference, and deployment stack for privacy-sensitive AI products.
NPM package DopplerBrowser-native WebGPU inference engine used across D4DA's local-first AI stack.
NPM package DoeZig-first WebGPU runtime for Node.js, Bun, Deno, and hardware-oriented execution paths.
Cerebras checks CSL receiptsBrowser-to-Cerebras verification receipts using Gemma 4 and Qwen evidence.
Public products
Legal review application built on D4DA's local-first AI runtime.
Disclosure control HarborPacket review workbench for policy findings, release checks, and evidence receipts.
Verified compute m3t4.aiPublic-facing arena and receipt surface for verified WebGPU compute.
Model behavior GammaBrowser game and comparison surface for model behavior experiments.
Agent system ReploidBrowser agent system with explicit state and tool use.
Data memory data.d4da.comDurable data comparison and memory surface.
Development cloud use
D4DA uses Google Cloud, Gemini, Claude, and ChatGPT for synthetic data generation, labeling assistance, evaluation, distillation, benchmark creation, and model training. D4DA products are designed to run local-first in privacy-sensitive deployments, with cloud inference available only when customer policy allows it.