Over the last year or so, LLMs have gotten shockingly good at reasoning, and enterprises aren’t just asking, “Can LLMs write SQL?” – they’re asking if LLMs can understand databases and design analytic pipelines?
LangGrant’s new LEDGE MCP Server is built to enable LLMs with an understanding of databases, and design analytic pipelines in response to user queries. The LEDGE server automates delivery of database context, without revealing any enterprise data, and enables the LLM to design analytic pipelines while complying with enterprise security and governance.
VMblog spoke with LangGrant co-founder, CEO and CTO Ramesh Parameswaran to dig into what LEDGE really does.
++
VMblog: To start, what does LangGrant actually do? What’s the core problem you’re solving?
Ramesh Parameswaran: The short version is LLMs are suddenly very capable, but enabling them to leverage enterprise databases is still a challenge. Databases are large and complex, so you can’t just feed the database to the LLM and hope for the best. Besides, asking LLMs to act on data is impractical for compliance, as most organizations need transparency and explainability in their use of AI.
At the same time, the demand for applying LLMs to enterprise databases is growing. Data teams want AI to help build pipelines, and automate engineering … but they need to apply AI in a way that’s safe and reliable.
So we designed LEDGE to first enable LLMs to understand databases by providing a complete database context that relies on metadata. No public LLM is exposed to enterprise data. Once the LLM has the context, it can produce multi-step analytical pipelines the same way a human would.
We also appreciate that LLMs need to interoperate with existing enterprise and cloud infrastructure and comply with existing data security policies and governance.
So that’s the core: help the LLM understand enterprise databases and design data solutions, while enhancing overall compliance and governance.
VMblog: The LEDGE MCP Server is your announcement. What problems is it really aimed at?
Parameswaran: Honestly, we continue to see a common pattern across enterprises.
Teams are trying to use LLMs for analytics, but they lack a robust system to provide the LLMs with sufficient context. People copy schema fragments into ChatGPT, provide a few example queries, and then hope the model “gets it.” It works sometimes, but it’s unreliable and manual.
And there’s the token mess. Because the model doesn’t understand the database, people throw more sample data at it – hundreds of thousands of rows in some cases – so the LLM can infer join patterns. It’s slow, expensive, and generates unreliable outcomes. It’s a pattern that reflects the widespread experimentation today. Enterprises have PII safeguards, authorization rules, residency constraints, firewalls, and a need for token governance – that can’t be bypassed.
LEDGE addresses these varied needs. LEDGE enables LLMs to comprehend databases with complete database context. No data leaves the governed system. Every interaction is governed according to enterprise access and security controls, and other governance requirements.
The end result is that teams get the full reasoning power of LLMs without the governance risks, manual effort, or explosion in token use.
VMblog: Could you walk us through a simple example of how LEDGE behaves?
Parameswaran: Sure. Let’s consider a simple user query:
“Show me total sales revenue for each product category last fiscal year.”
If you just ask an LLM this question without context, it’ll hallucinate or guess. With LEDGE the LLM understands the tables and relationships – the SalesOrderHeader, SalesOrderDetail, Product, and ProductCategory.
The LLM can reason:
“Okay, this joins through SalesOrderID, then through ProductID,” and so on. It builds the complete join graph for the query, and then specifies the time filter, groups by category, and the measures needed.
Even strong reasoning LLMs will occasionally hallucinate, and the LEDGE server validates LLM output to ensure the pipeline is accurate and executable.
The time required to deliver the data pipeline is usually about 1 minute. The pipeline is saved, easily reviewed and validated, and executed to deliver the data. The saved pipelines ensure that the use of the LLM is easily understandable and explainable, and transparent.
VMblog: How is LangGrant’s approach different from other AI workflow tools?
Parameswaran: A lot of AI-native platforms today expect you to build a semantic layer before AI can do anything. Ontologies, relationships, and curated metadata are all examples that can involve months of preparation and require ongoing maintenance.
LEDGE is different in 3 ways:
- LEDGE is a standalone automated database context engine.
No additional software, metadata curation, or knowledge-layer product is required, or ongoing curation of ontologies or other resources. - LEDGE LLM-generated pipelines are saved and orchestrated for repeatable, transparent, and explainable AI use
LEDGE saves LLM-generated pipelines that are easily reviewed and validated, modified, and run, addressing needs for ongoing AI transparency and explainability. - LEDGE is designed for limitless agentic development, with enterprise governance
LEDGE supports limitless agentic solution development and brings LLM-powered data under the enterprise data security and governance umbrella.
VMblog: As LEDGE enters Private Preview, what are you hoping to learn? And what challenges should enterprises expect?
Parameswaran: A lot of what we want to learn is really about scale in the wild. Not theoretical scale – real messy environments with Oracle and SQL Server and Postgres and Snowflake all living together. That’s where the rubber meets the road.
We’re also watching how teams plug LEDGE into their agent workflows. Everyone is experimenting with agents right now, so there’s a lot of variety in how people want these pieces to connect.
On challenges – yes, there’s always some initial hesitation. Even though LEDGE only exposes metadata, the idea of “an AI near the database” triggers some caution. That’s fair. Trust builds over time once people see how constrained and governed the system is.
And the other challenge is just mindset. Going from writing queries manually to letting an LLM generate multi-step pipelines feels different. Not harder – just new.
This whole category is brand new. A lot of what we’re doing in Private Preview is helping teams see where LEDGE fits alongside what they already have, and how to bring AI into analytics safely.
##
