By Most enterprise databases were designed for transactions and queries, not for AI agents that require contextual reasoning, scoped data access, and real-time governance. Bolting an LLM onto a fragmented, multi-system data environment without a purpose-built context layer exposes organizations to compliance failures, access control violations, and indefinite production delays. The gap between a working AI demo and a compliant, production-grade deployment routinely stretches to 18 months or more when teams attempt to close it with custom infrastructure.
In this interview on TFiR, Oren Eini, CEO and Founder at RavenDB, breaks down why enterprise AI initiatives stall before delivering ROI, how the inversion-of-control model in Quill eliminates LLM trust boundary risks, and what a real-world regulated deployment at scale actually looks like.
Guest: Oren Eini, CEO and Founder at RavenDB
Show: TFiR
Here is what every platform engineer, data architect, and enterprise AI practitioner needs to know.
Technical Deep Dive
Q: Why do enterprise AI initiatives stall between demo and production?
Oren Eini, CEO and Founder at RavenDB, points to a structural mismatch: enterprise systems were never designed, built, or conceived with AI in mind, so moving a working demo into production requires mapping every data source, addressing regulatory requirements such as SOC 2 and ISO compliance, and implementing access controls that prevent the AI from surfacing data it should not touch. The gap is not a technology problem in isolation. It is a systems architecture problem that typically requires near-total reconstruction of data pipelines. What looks like a few weeks of demo work becomes a year or more of infrastructure rebuilding before any user-facing value is delivered.
“All of the systems were not built, designed, or even conceived with the idea of AI. I have to rethink the whole thing.” — Oren Eini, CEO and Founder, RavenDB
Q: What specific data fragmentation problems block enterprise AI deployments?
In a typical financial institution, checking, savings, loans, mortgages, and credit cards each live in separate systems with separate schemas. Before an AI agent can reason across any of those systems, a team must unify the data, define the scope of access for each context, and transform the output into a form the agent can interpret. Eini describes this as open-heart surgery on a live environment. The engineering cost is not incidental. It is the primary reason AI projects fail to move beyond internal demos.
“You have to somehow bring them together, figure out the scope of access, and transform that into a form the AI agent can make sense of. That is akin to open-heart surgery in your environment.” — Oren Eini, CEO and Founder, RavenDB
Q: What is Quill and what problem does it solve for enterprise AI?
Quill is a context layer from RavenDB that sits on top of an existing SQL database. It reads the schema and structure of the database automatically, pulls the relevant data for any AI agent query, and manages the logistics of AI invocation and data access without requiring teams to rebuild their underlying systems. The intent is to compress what would otherwise be a 12 to 24 month infrastructure project into a deployment measured in weeks. Eini positions it as the missing layer between production data and production AI.
“Instead of going into this huge project where I am going to spend a year or two building that, the whole idea is that you roll off the production line within weeks.” — Oren Eini, CEO and Founder, RavenDB
Q: Why is MCP (Model Context Protocol) a security risk in public-facing AI agents?
MCP places the LLM in the position of deciding what queries to run and what data to fetch, which means the AI itself becomes the trusted actor in the data access chain. Eini argues this is appropriate for developer tooling such as Claude Code but is a structural security problem for any AI agent exposed to end users, because a sufficiently well-crafted prompt can cause the LLM to execute queries that surface data the user should never see. The entity in control of a production system should never be the AI model.
“If you are building an AI agent meant to be used by the public, MCP is a horrible idea because the entity in control is the AI, and the AI is not a trusted agent.” — Oren Eini, CEO and Founder, RavenDB
Q: How does Quill’s inversion-of-control model protect data in AI agent deployments?
Rather than allowing the LLM to generate and execute arbitrary queries, Quill defines a fixed set of named operations with explicit, predefined scopes. When an agent needs data, it calls a named operation such as “give me my credit card statements” rather than generating a dynamic SQL query with an LLM-controlled WHERE clause. The scope of what data is returned and under what conditions is determined entirely outside the model. This means the LLM functions as a natural language interface rather than an autonomous query engine, and the security boundary remains with the platform, not the AI.
“It is not ‘do a SELECT star FROM credit card WHERE id equals what’ where the entity in control of the ID is the LLM. It is ‘give me my credit card statements.’ Because the AI is not in control, you are in a much safer location.” — Oren Eini, CEO and Founder, RavenDB
Q: What is the right mental model for AI agents in enterprise systems?
Eini argues that most enterprises do not need general artificial intelligence. They need a reliable natural language interface into their existing systems. The practical need in most organizations is helping a user identify which form to complete, what to enter in a specific field, or how to navigate a process. Positioning the AI as a secretary rather than a super-intelligence changes both the security model and the scope of the engineering effort required to deploy it safely at scale.
“The point of the AI here is not to be the super-intelligence. It is to be the secretary, the natural language interface into my system. And that would be a huge thing to have.” — Oren Eini, CEO and Founder, RavenDB
Q: What industries and use cases is Quill best suited for?
Eini describes Quill as applicable across industries, with particular traction in environments where safety, compliance, and regulatory requirements make custom AI infrastructure prohibitively complex. A recent healthcare deployment serves as a concrete example: a fitness and activity tracking application that allows users to ask personalized health and exercise questions based on their profile. The platform handled the compliance and governance requirements, allowing the product team to focus on the application itself. The entire project, from specification through production, was delivered in under three months.
“The ability to not have to deal with all of that because this is already handled by the platform is a huge shift in the ability to deliver.” — Oren Eini, CEO and Founder, RavenDB
Q: What does the healthcare deployment case study tell us about Quill’s production readiness?
The healthcare application built on Quill is currently serving over 600,000 users in production. The full project timeline, including specification, was under three months. For a healthcare environment, this timeline is remarkable given the regulatory surface area involved. Eini uses this example to illustrate that Quill’s governance model is not theoretical. It has already absorbed the compliance burden for a regulated, large-scale, public-facing deployment.
“The entire project was under three months, including the specification, everything. And this is now in production, serving something over 600,000 people.” — Oren Eini, CEO and Founder, RavenDB
Q: What is the cost and resource comparison between building AI infrastructure in-house versus using Quill?
Building in-house requires committing three to five senior developers for approximately one year on the initial build, followed by another six months to a full year for deployment and stabilization. Eini estimates the total timeline at 18 to 24 months before a team has something stable in production. With Quill, onboarding is designed to be a guided, next-click process with AI-assisted setup, after which teams can begin iterating immediately. RavenDB offers the first three months free, with the expectation that teams will complete two or three full project iterations within that window.
“Doing it yourself means taking your three to five top developers and using them for a year. Doing it with Quill means plugging it into your system and clicking next a few times.” — Oren Eini, CEO and Founder, RavenDB
Resources and Documentation
- RavenDB, database platform and home of Quill, the AI context layer for enterprise SQL deployments
***
👇 Click to Read Full Raw Transcript
Swapnil Bhartiya: When it comes to databases, most enterprises build them around transactions and queries. Then came AI. And AI started demanding something completely different. Context reasoning, real time, data access, guardrails. Those things were not part of how you would build traditional database. So now organizations, teams are stuck. Either they migrate to new database for AI, which will take months, or, or they build a custom stack suitable for their database and AI needs, which once again will take months. So neither options are ideal. Now RevitDB says that there is a faster and better way. They just announced Quill, a context layer that sits on top of your existing SQL database and unlocks it for production AI agents, no rip and replace is needed. You will have full control. You can build guardrails, governance around that. And today we have with us Oren Eney, CEO and founder of revendb to walk us through why this really matters. First of all, Oren, great to have you on the show.
Oren Eini: Thank you for having me. It’s great being here.
Swapnil Bhartiya: It’s my pleasure. And as I was trying to lay the foundation of this decision, how correct or wrong I was when I was trying to explain when you look at the databases, how they were built and when you just plug in AI into it, what kind of friction, what kind of challenges organizations run into. So let’s talk about the problem area first.
Oren Eini: I think that the primary problem is the difference between the demo layer, the demo level, and actual production system. So let’s talk about the financial example. I go to my bank account, I download my credit card statement and I can go throw it at the chart and tell it how much did I spend on entertainment, how much did I spend on utilities. It will give me everything, including charts and information. Really nice. Now I go to my bank website and it’s pretty much unchanged from the 90s. So the key problem that we have to deal with is why is this the case? So in any modern company today, in any company that is, you know, bigger than mount up, you have the board saying, do AI, this is what we must do. And then the CEO, the CTO, CIO, whatever, goes to the IT department, sit with the R&D team, okay, do AI. And they take the three best people that they have and they do an amazing demo in a few weeks. And that’s really nice. And then you start looking at what does it take to actually take this demo and move it to production. Oh, you’re not a man and pop shop. You have to deal with regulation, SOC to ISO, et cetera, stuff like that. Well, what are the privacy that you have to Deal with how are you going to put guardrails in place so if someone asks nicely, they will be able to see your credit card details because you know the AI would do that for you if you’re not careful. So the whole idea here is that there is a huge gap in practice in implementing AI, typically in the order of I have to look into my system, map everything, almost rebuild it from scratch, because all of the system were not built, designed, or even conceived with the idea of AI. So I have to rethink the whole thing.
Swapnil Bhartiya: Thank you so much. Now, the thing is, as you mentioned earlier, that every company was like, hey, I want AI. And everybody is trying to get on AI bandwagon. But the fact is that you have seen that many enterprise AI initiatives, they kind of stall before they can actually deliver any. All right, the exciting part is, part is, sorry, let me just do it. I said far in the part. Like, as you rightly said, every enterprise wants to embrace AI and they are trying to, but in most cases, their AI initiatives stall before they can deliver any roi. So talk a bit about what does Quilt do differently to help them so that their AI initiatives, their AI projects are successful.
Oren Eini: So the chief problem is that the data that you have is not in one system. It is not envisioned to be used by AI. So in most cases, if you want to do AI properly, you have to first figure out what all the data is. And speaking of financial systems, you have check in, you have savings, you have loans, you have mortgages, you have credit cards. Each one of them is a separate system. Now you have to somehow bring them together, figure out what is the AA scope of access that you have to deal with that and transform that into a form that the AI agent can make sense of. That is the huge project that is akin to an open heart surgery in your environment. And what Quill allows us to do is just point our system into your database. It would understand the schema and the structure that you have, and it will pull the relevant details that you need on its own. So the whole idea is that we’re able to take care of all of the logistics of building the AI and invoking that and managing that for you without you having to deal with this. So instead of going into this huge project, I’m going to spend a year or two just building that. The whole idea is that you roll off the production line within weeks.
Swapnil Bhartiya: Can you also talk about, when we do talk about giving AI access to data databases, we have to also ensure guardrails, what they can do with the data or not. So can you also talk about how does Quill balance giving AI agents access to data but also ensuring governance is there that what they can or cannot do with the data.
Oren Eini: So let me try to talk a little bit for one second about MCP and model console protocol. This is a way for the AI to control, to make invocation, do calls, etc. And I think that if you’re building a tool like cloud code or something like that, it’s amazing. If you’re building an AI agent that is meant to be used by the public, it is a horrible idea because the whole idea is that the entity in control is the AI and the AI is not a trusted agent. The way that we looks at that, we invert the whole process. The editing control of what’s going on is. And Quil is very focused on insurance security. So when the AI needs data and when you ask a question about that, it is not oh, I’m going to invoke MCP and run some query and get the data for that because that’s a security enactment. Instead Quill defined well known queries and operations that you can execute and each one of those carry the scope that is involved with that. So it’s not do a select star from credit card where id what and the entity in control of what the ID here is the LLM. It’s give me my credit card statements. And that’s the effect that you have. Because the AI is not in control, you’re in a much safer location, you know what’s going to happen. And that is a huge difference. That inversion of control that okay, the point of the AI here is not to be the super intelligence, it is to be there to be the secretary, to be the natural language interface into my system. Because for most enterprises I don’t need a super intelligence. I don’t need just someone tell me what form I need to do and what I’m supposed to put in this location. That’s it. And that would be a huge thing to have.
Swapnil Bhartiya: Can you talk about what kinds of organizations or use cases is Quill best suited for right now? Or is it something that should become normal irrespective of what industry they are in?
Oren Eini: So I think that we are applicable in many different industries. We have really good success recently in a healthcare organization. And what they did, they built a go active sort of app and that helps you go with, you know, be more active, track your steps, encourage you on that, give you the ability to ask question. I’m so. And you know, I’m male, 40s, so and so diet. What should they do for exercise? It is able to give you some really concrete results. And if you think about that in a healthcare organization, the complexity that you deal with is again, safety, regulation, all of those sorts of things. And the ability to not have to deal with all of that because this is already handled by the platform, is a huge shift in the ability to deliver. The entire project, I think was under three months, including the specification, everything. And this is now in production, serving something over 600,000 people.
Swapnil Bhartiya: For those product leaders or CFOs who are watching this, what is the cost resource case for quilts versus doing it in house by themselves?
Oren Eini: Doing it in yourself is take your three to five top developers and use them for a year and then you have to deal with the actual deployment, which is another six months to another year. Doing that with Quill means plugging it to your system and click next a few times. And the entire thing is obviously we’re AI oriented and we’re doing the setup through the AI. So the entire thing is next, next, next, next, and you have something that you can immediately play with and start exploring. We are confident enough that what we have is that we are giving away the first three months. You know, go and do whatever you want. You get free access to everything. And by the end of three months, the expectation that you would be on your second or third iteration or project or new item that you can do in that time frame.
Swapnil Bhartiya: Oren, thank you so much for joining us and walking us through Quill and how it’s changing the way enterprises bring AI agents to their existing data. Because as AI is moving into production, this is a serious challenge. This is a serious problem and this is, this needs to be solved. So thank you so much for joining us and explain. And of course those who are watching, please go and check out Revendb.net and Quill for yourself how it can help you. And once again, thank you so much for chatting with me and I look forward to chat with you again.
Oren Eini: Thank you, thank you for having me.
