We build AI‑native products and explore what comes next.
Einstein Labs is an independent product studio and applied AI research lab. We build and run software for real problems, and we research the systems underneath: how to tell whether an agent did good work, and how to make agents reliable enough to trust.
Product Studio at the centre: Praxiom AI, agentFast, Propel.
NeoLab in orbit: 4 research bets on the systems those products run on. Hover a point to see what it is.
Product Studio
Products we build, launch and run.
Each product solves one problem for one audience, with its own site, roadmap and business model. Two are live today, and one is in development.
Praxiom AI
Product intelligenceProduct discovery from real evidence.
Praxiom turns customer interviews, support tickets, surveys and call recordings into scored insights, prioritised opportunities and PRDs that link back to the quotes behind them.
LiveVisit Praxiom AIagentFast
Agent infrastructureThe production framework for AI agents.
agentFast sits above whichever agent framework you already use and adds what production needs: durable execution, human approval before risky actions, memory, guardrails, evals, tracing and live streaming.
LiveVisit agentFastPropel
Growth agentYou ship. Propel sells.
Propel is a growth agent for technical founders. It reads your website and repository, works out who the product is for and what is worth saying, then runs marketing and sales as one loop.
In developmentAsk about early access
NeoLab
Research on the systems agents run on.
NeoLab turns open questions into experiments, prototypes and benchmarks. Every project carries a maturity label, and we say plainly when there are no results yet. None of these projects has reached the evaluated stage yet.
- Propel GymCan go-to-market agents be graded honestly, on their market beliefs, decisions and experimental discipline, inside simulated markets where the right answer is known?
- PMEvalCan an agent do the actual work of a product manager, and where does it fail?
- agent-streamWhat is the smallest event vocabulary that lets any agent stream a legible run to any client?
- PapyrusCan deterministic conversion keep enough structure and provenance for agents to cite their sources, without using a model to do it?
Research and products, one loop.
This is how we work, not a claim that every product uses every research result. Below are the connections that exist today.
- 1Research a hard problemNeoLab builds a way to measure it.
- 2Test it in a productReal users show what holds up.
- 3Bring back what brokeFailures become the next question.
Where the loop is real today
- Praxiom AIPMEvalMeasures the kind of work Praxiom automates.
- agentFastagent-streamagentFast streams its runs using this protocol.
- PropelPropel GymDesigned to measure Propel’s agent before any real market sees it.
Recent work
Dated, and limited to things that shipped.
Propel Gym: first answer-engine tasks built
The engine, the first two task types and the reference agents are built. Validity checks come next, before any results are published.
Papyrus released as open source
Document-to-Markdown conversion for agents, under the Apache-2.0 licence, with a live demo.
The production framework for AI agents went on sale as self-hosted, source-available code.
336 product management tasks across 14 categories, under the MIT licence. No models scored yet.
An event protocol for streaming agent runs, with Python and React reference implementations, under the MIT licence.
Read the research
The question, method and honest status of every research bet, with links to the open code.
Explore NeoLab