neob.dev is where NEOB does its thinking in public. It is not a product site and not a company brochure - those exist, and they are elsewhere. What is here is the research: the questions we are working on, the experiments we ran, the engineering decisions behind systems that are in production, and the parts that did not work.
What we work on
The unifying thesis is one agent, multiple interfaces, shared context. A business should not need separate AI systems for the phone, the website, a chat widget and a message thread. The channel is transport; what decides behaviour - knowledge, memory, business logic, tools, guardrails - belongs in one layer underneath all of them.
That thesis has consequences that most of the research areas are about: identity resolution across channels with very unequal evidence, latency inside a synchronous turn, tool interfaces that survive a probabilistic caller, and memory that can be deleted from exactly one place.
Three domains, three jobs
NEOB runs three public domains and keeps their roles separate on purpose - for readers, and so they are not competing with each other for the same search results.
| Domain | Role | Audience |
|---|---|---|
| neob.dev | Research, experiments, engineering write-ups, prototypes | Engineers, CTOs, technical founders |
| neob.ai | The company: AI solutions, consulting, delivery | Clients and partners |
| bitpull.ai | The product: an AI agent platform for customer communication | Businesses buying the product |
Content is not shared between them. This site is in English and technical; the product site is in German and commercial. If you are looking for pricing, industry pages or a demo, that is the product site. If you are looking for the company - NEOB Technology GmbH, its services and how to work with it - that is neob.ai.
How we publish
A few rules we hold ourselves to, because a research site that reads like marketing is not useful to anyone.
- Status is stated, not implied. Every experiment carries its real state: in production, prototype, exploratory or parked. Nothing that is not running somewhere real is described as shipped.
- Budgets are not measurements. Where a latency figure appears in a diagram it is an engineering target we designed against, labelled as such. We do not present design budgets as benchmark results.
- Limitations are mandatory. Every experiment has a limitations section. An experiment without one is an advertisement.
- Numbers we cannot defend are not published. Our evaluation sets are small, German-language and industry-specific. Where we cite a figure that belongs to a product - such as bitpull.ai's stated sub-500 ms response target - we say whose figure it is.
- Failures are published. The browser-agent work ended with us declining to make writes autonomous. That is on the site, in full, because it is the most useful thing we learned that year.
- Dates are visible. Publication and update dates appear on every article, so a reader can tell how old a claim is.
Who writes here
Pieces are published under two bylines. NEOB Research covers exploratory work and conceptual pieces; NEOB Engineering covers write-ups from systems that are running. Both are the same small team in Vienna, and the distinction is about the kind of evidence behind the piece rather than about who typed it.
NEOB Technology GmbH is a Vienna-based company; the legal detail is on the imprint.
Corrections and contact
If something here is wrong, we would rather know. Technical corrections, disagreements and questions: labs@neob.dev. Substantive corrections are made in the article with the update date changed, not quietly.
For commercial enquiries about the platform, use bitpull.ai; for company and project enquiries, neob.ai.