Kaarvi Founder Shravan Gunda on Goal-Based Agentic Analytics

Shravan Kumar Reddy Gunda ·  Founder & CEO, Kaarvi ·  Houston, Texas

MOST COMPANIES DO NOT HAVE A DATA SHORTAGE. THEY HAVE AN ACCESS PROBLEM. THEIR DATA IS SPREAD ACROSS FILES, DATABASES, CLOUD STORAGE AND APIS, AND GETTING A TRUSTWORTHY ANSWER STILL DEPENDS ON A SMALL GROUP OF SPECIALISTS AND A STACK OF DISCONNECTED TOOLS. GENERATIVE AI PROMISED TO CHANGE THAT. IN PRACTICE, MUCH OF WHAT SHIPPED WAS A CHAT BOX ON TOP OF THE SAME FRAGMENTED STACK: ONE QUESTION IN, ONE CHART OUT, AND A PERSON LEFT TO STITCH THE REST TOGETHER.

Shravan Kumar Reddy Gunda is the founder and CEO of Kaarvi, a Houston-based company building the Living Data Platform, an end-to-end agentic platform for data management and conversational analytics. Kaarvi is betting that the unit of work in enterprise data should not be a question or a prompt. It should be a goal. A user states the business outcome they need. The platform then surveys the data it can access, plans the work, asks for a single approval, carries out the plan, verifies the result and delivers it. The finished work is kept as a reusable starting point for the next time.

Before becoming Kaarvi’s CEO in January 2025, Shravan spent more than a decade in enterprise technology, most recently as an enterprise data architect in mortgage finance. Earlier, he was vice president of information technology at an AI company serving the oil and gas industry, and a senior system engineer at a global technology company. He holds a master’s degree in electrical and electronics engineering from Northwestern Polytechnical University.

Kaarvi operates with a three-person core team. This year it unveiled the Living Data Platform, moving its product from a traditional SaaS interface to a fully conversational, goal-driven experience. The company has filed more than 30 patent applications covering parts of the platform.

In this interview, Shravan explains what a goal-based approach means in practice, how it differs from a chatbot added to a business intelligence tool, the real-world problems it is designed to solve, and how a three-person team builds and reviews an enterprise-scale platform.

“Trust is the product. The goal is how you ask for it.” — Shravan Gunda, Founder & CEO, Kaarvi

Could you introduce yourself and tell us what you are building at Kaarvi?

I’m Shravan Gunda, founder and CEO of Kaarvi. We’re building a platform where you run your company’s data by telling it what you want to achieve, not by operating tools.

Before Kaarvi, I was an enterprise data architect in mortgage finance, and before that I led IT at an AI company serving oil and gas. In both roles I saw the same problem from two sides. In mortgage finance, loan data arrives from many lenders, and much of the work is proving that it is complete and consistent before anyone can rely on it. In oil and gas, the AI was only as good as the reservoir data underneath it. The analysis was never the hard part. Getting to data people trusted was. That is the gap Kaarvi is built to close.

The idea is simple. People who need something from data should not have to learn five tools, break their problem into a dozen technical steps, or wait in a queue. You connect your sources and state a goal in plain language. Kaarvi’s agents work out what data is needed, whether it can be trusted, what the plan is and what it will cost, and they ask you before anything runs. You approve once. They do the work, verify it, and hand you the result with a receipt showing exactly how it was produced.

Our tagline puts it plainly: “Talk to your data. Get it done.”

What is broken about the way companies work with data today?

Two things. The tools grew up in silos, and the burden of thinking still sits with the person.

The silo problem is well known. There is one product for ingestion, another for the catalog, another for quality, another for governance and another for BI. Each has its own interface and its own specialists. By the time a business user gets an answer, it has passed through several hands, and nobody can easily say where the number came from.

The second problem is newer and less discussed. Even with a good AI assistant, the person still has to break a business need into a series of technical questions: Which tables have this? Is the data complete? Join these two. Filter that. Now compare it to last quarter. The assistant answers each question well, but the person is still the project manager, the quality reviewer and the integrator. That isn’t self-service. It is a faster version of the same job.

You have described Kaarvi as “200% agentic,” and more recently as goal-based. What is the difference?

“200% agentic” was about the interface. The conversation is the product, not a feature added on. We started with a traditional SaaS interface, with a screen for every capability, and made the hard decision to rebuild everything so that agents could run it end to end through conversation.

Goal-based is the next step, and it changes the unit of work. A prompt asks for a single output. A goal describes an outcome. Take “I need to know why gross margin slipped in the Northeast region last quarter, and I need it in a deck for Thursday’s review.” That sentence is not a query. It is a small project. It involves discovery, quality checks, several queries, a comparison, a narrative and a deliverable. Goal-based means the platform owns the project, and you own the decisions.

Walk us through what actually happens when someone gives Kaarvi a goal.

First, it surveys. Before it promises anything, it reads which sources your organization has actually connected, such as files, databases, cloud storage and APIs, and checks whether the goal can be met from them. If something is missing, it says so up front and tells you what to connect, rather than producing a confident answer from half the data. We call this source awareness, and it is the step many agent products skip.

Second, it understands and plans. The platform has a large toolkit of skills, from profiling a dataset to detecting anomalies to building a dashboard. It selects the skills the goal requires and writes a plan you can read: these steps, in this order, using these sources, at roughly this cost. If the goal is ambiguous, it asks, but it keeps questions to a minimum.

Third, you approve. Nothing runs without your yes. For most goals, that is a single approval for the whole job. For recurring work, you can turn on autopilot, a standing approval that operates within the rules and spending cap you set. Even on autopilot, if the plan reaches an irreversible step, such as sending something outside the company, it pauses and asks again.

Fourth, it executes and verifies. The agents carry out the plan, and a separate verification pass checks the results and assigns a confidence level. The platform then delivers the result in the format you asked for, whether a deck, a dashboard or a report, along with a receipt: what was done, which sources were used, what was verified, how confident the platform is, and what it cost.

Finally, it remembers. The finished goal becomes a reusable base. Next quarter, you say “run that again,” and it re-plans against your sources as they are now, tells you what has changed since last time, and carries the lineage forward.

How is that different from a chatbot on top of a BI tool, or from the agents other vendors ship?

There are three differences, and none of them is the language model.

First, we plan against what you actually have. Many assistants assume the data is there and answer anyway. Ours checks first and refuses to pretend.

Second, the shape of control. You give one approval for the whole job, you see the cost before you say yes, and you can pause, stop or undo. Every action the platform takes on your behalf stays within limits you set. That is what makes it safe to hand it a project rather than a question.

Third, the receipt. In an enterprise, a chart with no history is a liability. Every result Kaarvi delivers comes with its lineage, its verification and its cost attached, and the workflow itself is saved so a colleague can rerun it or an auditor can review it. A person can check the work after the fact. That is a very different product from a chat window.

What real-world problems does this solve? Can you give some concrete examples?

Consider a manufacturer running a monthly margin review. Today, an analyst pulls data from the ERP system, two spreadsheets maintained by a plant manager and a pricing system, reconciles them by hand and builds the slides. With a goal-based platform, the operations lead states the goal once. The platform finds the four sources, flags that one plant’s file is missing last week’s numbers, and asks whether to proceed or wait. It then runs the comparison and delivers the deck with the caveat noted on the slide. Next month, it takes one sentence: “Run the margin review again.” The platform reports that the pricing table has gained two new columns since the last run and asks whether they matter.

Or picture a retail or sales organization trying to understand why a region’s numbers dropped. The goal touches orders, returns, promotions and staffing. The plan lays out each step. The verification pass shows which conclusions are well supported and which rest on thin data. The regional manager gets an answer they can defend, and the analyst who used to assemble it gets their week back.

In oil and gas, much of the pain comes from data quality across systems that were never designed to work together. A goal like “Give me a trustworthy production summary for these assets” is mostly a data engineering problem before it becomes an analytics problem. The platform can profile the sources, run quality checks, propose fixes and ask for approval before changing anything. That last part matters. Nothing gets “cleaned” silently.

In finance, the pattern is reconciliation and the period-end close. The value is not the chart at the end. It is that every number in the chart has a lineage you can hand to an auditor.

The same approach works beyond analytics. Our AI-powered personalized outreach workflow tracks dark traffic, the anonymous research buyers do before they ever fill in a form, and pinpoints likely buyers with 95%+ accuracy. We measured that against 60 buyers over three months.

How do you stop an AI agent from giving a confident but wrong answer on enterprise data?

This is the question that matters most, and I would rather be honest than claim it is solved. Language models can be wrong, and in analytics, a wrong number delivered confidently is worse than no number at all.

The goal-based design is part of the answer. The platform checks what data exists before it plans. The plan is visible before it runs. The results go through a verification pass that states a confidence level instead of simply asserting an answer. The receipt shows the sources and the steps. And a person approves before anything changes.

Underneath that, the agents work from real structure instead of guessing. Lineage shows where data came from and what depends on it. Profiling and quality checks show whether the data can be trusted before anyone queries it. Governance and classification determine what an agent is allowed to touch, and role-based access control sits on top.

Every answer shows its work, and every number traces back to the query that produced it. I made this case in a recent piece for MSN, “Your AI sounds confident. Ask it for a receipt.” A confident answer you can’t check doesn’t belong on a board slide.

Our internal engineering standard is written down plainly: no assumptions, no hallucinations, and every claim backed by a source. We hold how we build to the same bar as what we sell.

Enterprise AI can get expensive quickly. How do you think about cost?

Cost has to be visible before the work starts, and it should be tied to the goal, not to the number of prompts someone typed. We use credit-based billing. The plan you approve includes a cost projection, and autopilot runs within a cap you set. A simple lookup does not cost what a multi-step goal costs, and you see the difference before you say yes.

Consolidation is the other lever. When ingestion, cataloging, quality, governance, pipelines and analytics run in one loop, you aren’t paying for, integrating and staffing a separate tool for each step.

Kaarvi runs with a three-person core team. How do you build a platform of this scope?

We run the company the way the product runs a goal. AI agents are part of how we build, but we treat them the way a good engineering manager treats a fast new hire: clear rules, clear boundaries and a lot of review.

We follow a written process that runs from planning through testing, verification and deployment, and ends with a review of what we learned. Agents write code and deploy to our test environment. Production is owned by people, and agents never touch it on their own. It is the same principle the product gives customers: nothing irreversible happens without a person’s yes.

The part I’m proudest of is review. We run several AI review agents in parallel, each with a different job. One acts as a security adversary. One looks at scale and query plans. One hunts for silent failures, the bugs that don’t crash but quietly return the wrong result. One compares what was planned with what was delivered. A small team can hold a high bar when every change is examined from four angles.

We even use Kaarvi to market Kaarvi. Our product videos come from a workflow inside the platform called VideoBox. It reads the product’s own code, writes the script, generates the visuals and voiceover, and checks every claim in the script against what the code actually supports, so a video can’t say something the product doesn’t do. For a team our size, that replaces an agency.

Kaarvi has filed more than 30 patent applications. Where is the real innovation?

The individual pieces, such as a catalog, a quality check or a SQL generator, already exist in the market. The hard part is getting agents to take a stated goal, plan it against the data that actually exists, run it within a person’s limits, verify it, and hand back work that someone can check and rerun.

Approval gates and audit logs are quickly becoming standard in agentic AI, so that is not where we claim anything new. Our filings cover how the pieces work together: checking a goal against the data a company has actually connected before any plan is made, tying a single approval to a costed, multi-step plan across data management and analytics, and attaching a receipt that allows the same goal to be verified and rerun later.

How should companies measure whether agentic analytics is working?

Start with the goal, not the model. Measure how long it takes to get from a stated business goal to an outcome someone trusts enough to act on. Measure how many people can set that goal themselves without filing a ticket. And measure how many goals run a second time without anyone rebuilding them.

Those three numbers, time to a trusted outcome, the share of people who can self-serve, and how much work compounds, tell you whether the platform is changing how the business runs. Model accuracy matters, but it is a means to that end.

What will define the next phase of AI in enterprise data?

The shift from answers to outcomes. Today, most tools stop at “Here’s the chart.” The next step is “Here’s the goal you set, here’s the plan. Approve it, and here’s the verified result with its receipt.” And the step after that is “Here’s what changed since last time, and here’s what I’d do about it, if you agree.”

The companies that win won’t be the ones with the flashiest demo. They will be the ones whose agents are governed, auditable and affordable enough to run every day. Trust is the product. The goal is how you ask for it.