Shamrock AI

Future AGI, founded in 2024 in San Francisco, provides tools for creating synthetic data, testing AI models, analyzing errors, and improving AI performance.

Pre-Seed · $2M · 2025-02

Arga Labs provides high-fidelity sandbox environments for testing and training AI agents, building stateful service twins that clone the APIs, CLIs and MCPs of third-party enterprise software such as Salesforce, Workday, Stripe, Slack and Google Drive, with permission systems and webhooks left intact. Developers point a single API base URL at Arga's twins to run agents against a replica instead of production, seed scenarios with mock or production data, run automated evaluations against defined criteria, and trace every action. It targets mid-to-large engineering teams building their own AI agents or deploying third-party autonomous workflows, solving the problem that agents cannot safely fail in live enterprise systems and that production software cannot be reset repeatedly for reinforcement-learning training.

Seed · 2026-08

Patronus AI, founded in 2023 in San Francisco, USA, builds large-scale simulation environments mimicking websites and internal systems to conduct stress testing and reliability evaluations of AI agents before deployment.

Series B · $70M · 2026-06

Colossal-AI, founded in 2021 in Beijing, is a developer of high-performance AI runtime platforms.

Series B · $138M · 2026-03

Langfuse is an open-source AI engineering platform for LLM observability, evaluation, prompt management and experimentation. Founded in 2022 and backed by Y Combinator (W23), it is based in Berlin with an office in San Francisco; customers include Merck, Canva and Twilio. ClickHouse acquired Langfuse in January 2026, and the product remains MIT-licensed open source.

$4M

Inephany, founded in 2024 in London, UK, develops AI-driven software that optimizes training and deployment of large-scale machine learning models, initially focusing on training-time optimization for LLMs.

Undisclosed · $2M · 2025-04

Arato's platform learns the behavioral patterns of enterprise AI applications and automatically generates tens of thousands of test scenarios simulating real human behavior — interacting via text, voice, images, and product interfaces — to surface potential failures, biases, and risks before deployment without requiring code access, and monitors performance post-launch. The company was co-founded in 2024 by two founders of Stoke Talent, which was acquired by Fiverr for $95 million in 2021.

Seed · $10M · 2026-06

Datumo, founded in 2018 in Seoul, South Korea, provides a large language model evaluation and benchmarking platform. Datumo stands out through its licensed datasets, particularly data scraped from published books.

Undisclosed · $16M · 2025-08

Adansons, founded in June 2019 as a Tohoku University spinout building on Prof. Yoshitaka Kimura's reference-based AI theory originally developed for fetal ECG extraction, offers MLDebugger / ML Atlas, a SaaS platform that makes ML model inference in autonomous driving and similar systems observable, debuggable and reliable, serving automotive OEMs and Tier-1 suppliers.