Vaibhav Gupta, CIO, IndiaMart discusses AI-powered search, voice automation, buyer-seller matchmaking, semantic data layers, AI accountability, and why the CIO role is evolving from information management to intelligence and innovation.
With millions of interactions taking place across thousands of categories, languages, and business requirements, a platform must continuously understand user intent, establish trust, and deliver relevant matches in real time. As AI becomes increasingly embedded into marketplace operations, it is reshaping how discovery, engagement, and customer interactions are managed at scale.
Vaibhav Gupta, CIO of Indiamart, shares how the company is leveraging AI across search, buyer-seller matchmaking, voice automation, and business intelligence to power the B2B marketplaces. He discusses the evolution of Indiamart’s proprietary AI models, the architecture behind its voice AI platform IM Vani, the importance of semantic data layers, and the company’s approach to balancing automation with accountability.
CIO&Leader: Indiamart serves millions of users across India and caters to an extremely diverse customer base. How do your AI systems solve the intent-versus-language challenge?
Vaibhav Gupta: One of the unique aspects of India is its linguistic diversity. The country encompasses more languages and dialects than many continents combined. Building a marketplace that effectively serves SMEs across such diversity has always been a significant challenge.
To provide some context, Indiamart operates across nearly 98,000 categories. That is roughly five times larger than the category breadth of a typical e-commerce platform. Large e-commerce companies may operate around 20,000 categories, whereas our depth and breadth are substantially larger.
Customers across geographies often search for the same product using different languages, dialects, spellings, or even images. Addressing this challenge has been a long journey.
Long before the current wave of generative AI, we collaborated with several academic institutions nearly a decade ago to develop our own models for intent identification. These models were designed to understand what a user is actually searching for, regardless of whether the query contains spelling errors, regional language variations, or image-based inputs.
The objective is to map every query to the correct intent and then connect that intent to the relevant product, seller, or category in our database.
The AI models also help us benchmark semantic similarity. For example, if a user searches for the same product using different spellings or linguistic variations, the system must still arrive at the same result.
Once the product intent is correctly identified, the next challenge is matching the buyer with the most relevant seller. That process is powered by an internally developed AI-based matchmaking engine. After identifying the product, the system determines which sellers carry that product and which categories are associated with it.
These models were built well before the emergence of generative AI. They are proprietary AI systems developed specifically for search, intent recognition, and buyer-seller matchmaking.
CIO&Leader: As AI agents become increasingly autonomous, how do you manage the risk of hallucinations or incorrect commitments during buyer-seller interactions?
Vaibhav Gupta: We did not replace our chatbot with a voice agent. Chat and voice are separate interaction channels, and our objective is to automate both.
The opportunity emerged from a recurring business scenario. Buyers frequently attempted to contact sellers, but sellers sometimes missed those calls. In such cases, the calls would be routed to Indiamart’s call center.
By the time the call reaches us, we already know which product the buyer was viewing. The conversation therefore focuses on confirming requirements and collecting relevant details.
For example, if a buyer is searching for a particular industrial product, the system may need information such as specifications, quantity requirements, budget, and location. We built an AI layer capable of determining which questions should be asked based on the context.
That intelligence layer focuses on identifying relevant questions around specifications, quantity, price, and location. On top of that layer, we added a voice layer to facilitate communication.
The architecture is designed to minimize hallucinations because it is not a single monolithic AI system. Multiple layers work together.
The question-generation layer is built and controlled by Indiamart. The voice layer is provided by an external partner. Each layer is independently designed and evaluated. Multiple agents operate across different layers to ensure reliability.
Most hallucination risks arise within the conversational layer. The underlying question-generation framework contains structured logic and predefined parameters, which significantly improves accuracy.
CIO&Leader: Indiamart has moved from keyword-based search toward behavioral lead matchmaking. How do you balance product relevance and seller trust in real time?
Vaibhav Gupta: Seller verification and behavioral matchmaking are separate systems.
Seller verification operates independently of lead matchmaking and is designed to establish trust within the marketplace.
Every seller must go through verification processes. The level of verification depends on the type of seller account.
For free sellers engaging with buyers on the platform, GST verification or document-based verification is mandatory.
For sellers enrolled in paid packages and trust programs, we conduct additional checks including video KYC verification and several other validation mechanisms.
Trust determines marketplace eligibility; AI determines marketplace relevance
This trust framework determines whether a seller can participate on the platform and helps prevent unverified sellers from accessing buyer leads.
Behavioral matchmaking begins after that trust layer has been established.
Once buyers start interacting with verified sellers, we analyze behavioral signals such as response rates, category engagement, geographic preferences, purchasing patterns, and transaction volumes.
Based on those behavioral signals, the platform identifies the most relevant buyer-seller matches and recommends leads accordingly.
Trust determines marketplace eligibility. Behavioral matchmaking determines marketplace relevance.
CIO&Leader: For a platform processing millions of interactions, how do you manage latency, disaster recovery, and AI workloads across cloud environments?
Vaibhav Gupta: To improve latency, we migrated our cloud infrastructure from the United States to India. Even small improvements measured in milliseconds can significantly impact customer experience at our scale.
Although we operate on a dual-cloud architecture, different workloads are assigned to different cloud environments.
Marketplace traffic is hosted on the cloud environment optimized for low-latency customer interactions. Business intelligence and analytics workloads operate on a separate cloud infrastructure.
Latency-sensitive applications and data-intensive analytical workloads have very different requirements, and separating them allows us to optimize both.
For AI systems, we rely heavily on a semantic data layer built internally by Indiamart. This semantic layer resides within our data cloud environment and provides structured context for AI-driven applications.
Most of our current generative AI applications do not require ultra-low-latency, real-time database lookups.
Even in voice interactions, there is typically a three-to-four-second call connection window, which provides sufficient time to retrieve necessary information.
The more demanding scenario occurs when an AI system must process a query, retrieve information from a database in real time, and return a response within milliseconds.
To prepare for that future, we are building a semantic layer that identifies important business metrics, attributes, and conversation elements in advance.
This layer is being built on Redis, allowing agents to retrieve both data and contextual meaning in approximately ten milliseconds.
We are already using this capability for voice interactions as well as email and WhatsApp-based conversations, and it will become increasingly important as real-time AI experiences evolve.
CIO&Leader: As AI agents become more capable, how do you define accountability when an autonomous system makes a mistake?
Vaibhav Gupta: One reality of AI is that hallucinations and errors are possible. Organizations cannot avoid that fact.
At the same time, responsibility cannot be shifted to the AI system itself.
AI is performing work on behalf of the organization, much like any employee would. If an employee makes a mistake, leadership remains accountable. The same principle applies to AI systems.
AI is like any other employee, the CIO remains accountable for its failures.
The responsibility ultimately rests with the CIO and the teams building and operating these systems.
That is why we focus heavily on creating ensembles of algorithms, monitoring mechanisms, and analytics frameworks that reduce failure probabilities and identify issues quickly.
A good example is our customer care email automation initiative. Approximately 85% of incoming customer emails are now automated.
During initial deployment using a single algorithm, we observed nearly 10% inaccuracy rates. That level of error was unacceptable for customer communication.
We therefore introduced ensemble approaches. When multiple systems disagree on an answer, those cases are routed to human reviewers instead of being automatically processed.
Could multiple systems still produce the same incorrect response? Yes, although the probability becomes extremely low.
For those scenarios, human monitoring and analytics functions remain in place to identify and correct issues.
Accountability ultimately remains with the organization, not the AI.
CIO&Leader: Where does human-in-the-loop fit into your AI strategy?
Vaibhav Gupta:Human involvement depends entirely on the use case.
Some AI workflows are fully automated. Vani’s lead-handling conversations are examples where automation can operate independently without requiring human intervention.
Other use cases are recommendation-oriented. In those scenarios, AI generates intelligence, identifies clusters, segments users, and recommends actions.
A third category involves human-in-the-loop workflows. For example, we have systems that assist with meeting scheduling and productivity enhancement. In these cases, AI performs preliminary tasks before handing the process to a human operator.
Different combinations of automation, recommendations, and human oversight are appropriate depending on business requirements, return on investment, and accuracy expectations.
CIO&Leader: Looking ahead, how do you see the CIO role evolving over the next few years? Will the “I” in CIO continue to stand for Information?
Vaibhav Gupta: Historically, the “I” represented Information because technology teams were primarily responsible for providing data, reports, forms, and ERP systems.
That role has already evolved significantly.
Today, technology organizations are increasingly responsible for delivering business intelligence rather than simply information. Over time, CIOs will become even more accountable for business KPIs themselves.
The role is moving from information toward intelligence and innovation.
Future CIOs will focus less on managing systems and more on customer experience, business strategy, productivity, technology-enabled growth, and measurable business outcomes.
The future CIO will own business intelligence, innovation, and customer experience
The objective is no longer simply maintaining infrastructure. It is enabling and managing the performance of the entire business ecosystem.
That is why I believe the “I” in CIO is increasingly shifting from Information toward Intelligence and Innovation.