Einstein Labs is an independent AI product company and applied research lab. It has two parts. The Product Studio builds, launches and runs AI-native software products. NeoLab researches the systems those products depend on: how to tell whether an agent did good work, and how to keep agents reliable in production.
We do not sell client services. Everything on this site is either a product you can use, research you can read, or code you can run.
The products
- Praxiom AI turns customer interviews, support tickets and other research into insights and PRDs that link back to the quotes behind them. It is live.
- agentFast is a production framework for AI agents: durable execution, human approval, memory, guardrails, evals, tracing and streaming above whichever agent framework you already use. It is live.
- Propel is a growth agent for technical founders. It reads your website and repository and runs marketing and sales as one loop, saying only what it can prove. It is in development.
The research
NeoLab works on one primary question: how do you grade an agent’s work honestly when real outcomes are slow, noisy or partly subjective? Its secondary track is what an agent needs to survive production. Two further directions, context and memory and efficient inference, are exploratory.
- Propel Gym: simulated markets where go-to-market agents are graded against a hidden truth. In progress.
- PMEval: a 336-task benchmark for agents doing product management work. Prototype; no models scored yet.
- agent-stream: an open event protocol for streaming agent runs. Prototype; used by agentFast.
- Papyrus: document-to-Markdown conversion that keeps page anchors so agents can cite sources. Prototype.
Why both
Products are where AI meets real users, real data and real bills, so they show quickly which capabilities hold up. Some of what products need does not exist yet, such as an honest way to grade an agent’s marketing decisions. NeoLab works on those problems directly. Research creates capabilities, products test them in real use, and what breaks in real use becomes the next research question.
Not every product uses every research result. Today there are three real connections: Propel Gym was designed to measure Propel’s agent, agentFast streams its runs with the agent-stream protocol, and PMEval measures the kind of work Praxiom automates.
How to read our status labels
Every research project carries one of five labels: proposed, in progress, prototype, evaluated or published. A project moves up only when the evidence does. As of this post, none of our research projects has reached the evaluated stage, and the NeoLab page says so. We will publish negative and inconclusive results alongside positive ones.