Adaptiwisdom
Embodied AI · Agent infrastructure

AI you can measure before you trust it.

Adaptiwisdom is an independent technology company building measurement-first AI infrastructure: data-quality tooling for robot learning, a governed control plane for enterprise AI agents, and specifications that coding agents can build from.

See what's shipping contact@adaptiwisdom.com

Flagship · Embodied intelligence

Pier: better robot models start with measured data.

Pier is an embodied-intelligence platform: teleoperation data from a robot fleet, trained into vision-language-action (VLA) models and deployed back to the robots. Today's focus is the part the field under-measures, the quality of robot-learning data itself.

In active development

Pier data-quality toolkit

Open, reproducible measurement of robot demonstration data

A command-line toolkit that scans LeRobot-format datasets (SO-100/SO-101 arms and LIBERO) and scores every trajectory on ten quality axes: smoothness, action–state alignment, redundancy, diversity, gripper chatter, frozen or stuck channels, operator handovers and more. It runs offline on a CPU, and its reports are reproducible byte-for-byte.

10quality axes per trajectory
50+open datasets measured
$0to reproduce, CPU-only
pier scanpier qualitypier curatepier judge-auditpier screen

Latest

  • DQ-Bench v0.1: an open robot-data-quality benchmark anyone can reproduce at no cost.

  • External validation on labelled failure datasets (RoboVAD, BotFails, ArmnetBench): provenance signals separate policy rollouts from human teleoperation at AUC 0.89–0.97.

  • New curation axes (frozen channels, alignment lag, stuck-vs-command) validated on 43 real datasets; support for LeRobot v2.x datasets.

  • Operator handovers measured across 3,296 switches on 50 public datasets.

  • Quality Index v0.2.0: a scored audit of the open SO-10x data commons on a registered sample of 100 datasets.

Next: real-fleet data collection, GPU training of VLA models and safe over-the-air deployment, phased in as resources come online.

Products

Infrastructure for agents that act in the world.

Production-ready

Phan

Governed control plane for enterprise AI agents

Create AI agents, attach approved tools and run auditable workflows, with human approval for risky actions and full cost control.

  • Human-in-the-loop approvals and allow / deny / require-approval policies
  • Tamper-evident, hash-chained audit log with signed evidence bundles
  • Durable workflows, per-run cost tracking and budget caps
  • Multi-provider model routing, OIDC SSO and tenant isolation
In development

Author

From a plain-language requirement to an agent-ready spec

A command-line tool that turns a description of a system into a complete design package (architecture, decision records, interface specs and a task graph with testable acceptance criteria) for coding agents to build and maintain.

  • Every requirement traced to a task and an executable check
  • Documented, agent-agnostic package format
  • Evolves the package as requirements change

Approach

Measurement before claims.

Pre-registered studies

We state predictions before testing, and keep negative results on the record next to positive ones.

Reproducible by default

Deterministic pipelines, pinned fixtures and outputs checked byte-for-byte in CI.

Open collaboration

Built with a distributed network of engineers from the open-source community.

Contact

Partnerships, pilots, or research collaboration.