Saran Kumar Krishnasamy spent close to six years at Dataminr building real-time NLP systems that processed petabyte-scale data and helping grow its AI organization to fifty researchers and engineers. He left that staff-level role in early 2024 to co-found Gigit.ai with one bet: consumer product discovery is moving from Google’s search box to ChatGPT, Perplexity, and Google’s AI Overviews, and the e-commerce brands that don’t restructure for AI crawlers will quietly disappear from the answers shoppers see.
Today, Gigit.ai is live with more than twenty merchants, has hit $100K ARR within four months of launch, and has raised $870K from Peak XV Partners, Pear VC, and Forge Ventures. Saran holds an M.Eng. from Cornell and is a published researcher in LLM evaluation at Eval4NLP 2023 (AACL-IJCNLP). He’s one of a small group of technical founders defining Generative Engine Optimization as a category.
What pulled you toward the U.S. specifically, and how did that transition shape the way you think about building technology?
I grew up in India and moved to Singapore for my undergrad at NUS, where I got a solid technical foundation. But 2016-2017, the U.S. was the place where AI was translating into real companies at an extraordinary pace. Uber was already valued in the tens of billions, Robinhood was on its way to unicorn status, Netflix had shown that recommendation systems could define an entire consumer experience, and a new billion-dollar startup was minted roughly every week. AI wasn’t a theory anymore, instead, companies were being built on it, and consumers were already changing how they discovered and bought things.
I came to grad school in 2017 at Cornell Tech, where the program was a mix of CS and entrepreneurship. I co-founded a fintech startup that leverages data and financial services with peers and we were finalists for the startup awards. My AI research there led to joining Dataminr in 2018 as one of the early AI research engineers. The startup ambition and the deep technical work were running in parallel from the start. GigitAI is where those two threads finally came together.
Your early career took you through PayPal and Visa, where you worked on payment processing infrastructure serving millions of transactions. What did building systems at that scale teach you about how commerce actually works at the foundational level?
My first real industry experience was an internship at PayPal during my third year at NUS. My first full-time job out of college was at Visa working on dispute resolution systems, building backend services and large-scale batch processing pipelines for three years. I learned a lot there and grew so much as an engineer.
What I remember most is how seriously Visa took reliability at scale. You couldn’t afford to have things break when you were processing millions of transactions, so the bar for any change going into production was high. In 2015-2016, the industry was buzzing about Kafka, Cassandra, Spark and startups were building entire pitches around these tools. Visa’s approach was more measured: validate, test at scale, and then roll out. That made me respect the infrastructure layer. We work at start-up speed at Gigit, but when a merchant hands over their storefront and revenues to you, the systems behind the scenes have to be rock solid. The instinct to get that foundation right, before building fast on top of it, came out of Visa directly.
At Dataminr, you went from engineer to staff-level leader, helping grow the AI team from a small group to 50 researchers and engineers while building NLP systems that processed petabyte-scale data in real time. What was the hardest part of that scaling process, and what did it reveal about how AI teams succeed or fail?
The biggest lesson learned from that scale process was the importance of the composition of the team. Dataminr’s AI leadership was deliberate in forming multidisciplinary teams: research scientists in NLP, computer vision, knowledge graphs, and HCI working alongside engineers and product experts. This diversity of background kept the work from becoming abstract.
But the real problem as the team grew was the gap between research and production. The number of AI models being developed kept growing, but the infrastructure to productionize them couldn’t keep pace. It’s the classic research-to-production gap that MLOps was built to address.
I was on the AI engineering team that created a framework to bridge that gap. The goal was to empower researchers themselves to deploy their models for real-time use cases, without having to have a separate engineering team to translate everything. And that changed the whole speed of the organization.
What I learned from this is that AI teams don’t fail because the models aren’t good enough. They fail because the distance from research to production is too long or too painful. We built GigitAI with that in mind from day one.
You published research on using smaller LLMs as evaluation metrics for summarization at the Eval4NLP workshop, co-located with AACL-IJCNLP. That’s a fairly niche technical contribution. How does that research connect to what you’re building now at Gigit.ai, if at all?
That paper was a collaboration with colleagues and focused on the use of smaller language models as evaluation metrics for summarization. Everyone was deploying the biggest models they could get their hands on at the time, but real-time systems couldn’t take advantage of them due to latency constraints. The paper explored one angle of that problem, but more broadly, working on real-time AI systems meant we were always thinking optimization: hardware accelerators, quantization, distillation, all the techniques that let you get meaningful performance out of smaller, faster models. This experience taught me to always think about the trade-offs between model capability and deployment constraints. That same thinking runs through GigitAI; we personalize storefronts in real time on every visit, which means our models have to be both fast and accurate. There simply is no version of this where you trade one for the other.
You left a staff engineer role at a well-funded company to co-found Gigit.ai and tackle Generative Engine Optimization. Walk us through the moment or observation that convinced you this was a big enough problem to bet your career on.
By the close of 2023, LLMs had reached the point where applications we’d only theorized about suddenly were buildable. I’d spent close to six years building real-time AI systems at Dataminr and had spearheaded the development of an internal LLM. It felt like everything I’d learned over the past decade was culminating in this moment where the technology was finally ready for a new class of products. I’d been thinking about starting something for a while and this felt like the right time.
My co-founder Inez and I have known each other since our undergraduate days at NUS. We had been meeting regularly to brainstorm ideas and in early 2024 we decided to take the plunge. We began with an AI-powered customer support bot and launched with a handful of merchants. In the process, we began to experiment with displaying personalized segments of text on product pages, depending on how the visitor interacted with the site. That’s what opened the door to real-time personalization of storefronts.
Once we saw what was possible there, the GEO connection was obvious. What turns a human shopper is exactly what AI engines need to find and recommend products, the same structured and dynamic content. Product discovery was shifting from search bars to AI-generated recommendations, and there was no infrastructure to help smaller brands participate in that shift. That was a big enough problem to bet on.
While the majority of people still think of SEO when they think about online product discovery, we’re seeing that the entire model is rapidly change because consumers are finding products through ChatGPT Google’s AI Overviews, etc. rather than traditional search results. What does that shift actually look like in the data you’re seeing from your merchants?
When we look at merchant analytics, AI-referred traffic from places like ChatGPT still makes up a small percentage of overall visits. But it’s growing and the behavior is more interesting. Visitors coming through AI engines are significantly higher intent. They’ve already told an AI assistant what they want, given their requirements and been shown a product that fits. By the time they reach the storefront, they are no longer browsing. They’re evaluating.
That’s a very different kind of visitor than someone who clicks through Google search results or scrolls through a category page. And that changes what the storefront needs to do. With these kinds of visitors, you are working against yourself if you build a generic product page trying to appeal to everyone. They came with a specific need and expect the page to address it. That’s precisely the problem dynamic storefronts solve – matching the content to the visitor’s intent in real time.
Direction is clear, but shift is early. The merchants that are set up for this kind of traffic will convert it. The ones that don’t will lose high-intent shoppers to their competitors that do.
Gigit.ai builds dynamic storefronts that are designed to be read by both human shoppers and AI crawlers simultaneously. That’s a genuinely unusual technical challenge. Can you explain the architecture behind that without getting too deep into the weeds, and where the hard engineering problems live?
Most storefronts are built for one audience or the other but we built for both.
We start by ingesting the merchant’s product catalog, brand content, external domain sources like medical journals for health brands, and real-time signals like ads context, CRM data, and on-site behavior. From that, we build a knowledge graph that structures product information as entities and relationships. A beard oil isn’t just a SKU with a title and price. It connects to skin types, grooming goals, ingredients, and seasonal conditions. That knowledge graph becomes the single source of truth for everything downstream.
On top of it, we run a multi-agent content pipeline where specialized agents handle attribute extraction, intent classification, content curation, and content generation, all governed by a brand voice enforcement layer. The output splits into two rendering paths: a personalized storefront assembled in real time for human visitors, and a stable, schema-marked canonical layer for AI crawlers and agents.
The hardest engineering problems are at the boundaries. Keeping the knowledge graph accurate as merchants change their catalogs in real time. Making sure the multi-agent pipeline generates content that’s on-brand and factually correct, not just plausible. And maintaining sub-second response times for the personalization layer while serving stable, crawlable content on the same pages.
For a pre-seed company, you’ve had some rapid traction. What do you think those early signals say about the market timing for this category?
We’re live with 20+ merchants, primarily health, beauty and wellness. What really gets me is not just the number, but how fast merchants want to do something once they see the problem. This is not a long education cycle. When you show a merchant their products are invisible to AI engines or that every visitor sees the same generic storefront regardless of intent, the reaction is immediate.
The urgency tells me it is the right time. We’re not selling a solution to a future problem. Merchants know that the transition from traditional search to AI-driven discovery is underway. The ones we’re working with aren’t risk-taking early adopters. They are business owners who are watching their existing channels change and want to do something before they’re left behind.
A lot of your work is focused specifically on helping small and mid-sized U.S. retailers compete against larger players who have more resources to adapt to AI-driven commerce. Why does that segment of the market matter to you personally, and what happens to those businesses if they don’t adapt?
I’ve always believed in democratization and open access. Open-source shaped how I think about building technology: the best tools should be available to everyone, not just the people who can afford to build them from scratch. I carry that same lens into how we think about GigitAI.
Our earliest customers, who are still some of our strongest supporters, would tell us about features they saw on Amazon and wish they could offer their own shoppers. Things like Rufus, Amazon’s AI shopping assistant, or the personalized recommendations that adapt to every visitor. They’d say, ‘that would be incredible on our site,’ but they simply didn’t have the engineering team or the budget to build anything like it. These are merchants who care deeply about their customers and their products, but they’re locked out of the tools that would let them compete.
That’s what drives me. If the best AI-powered shopping experiences only exist on Amazon, then AI just becomes another force for consolidation. The merchants who can’t access those capabilities don’t just lose market share, they become invisible.
