Hiring a delivery worker in rural Bihar is nothing like filling a corporate role in Bengaluru. The candidate may speak three dialects, own no smartphone, and drop a call twice before finishing a sentence. Yet Vahan.ai, under the leadership of Founder and CEO Madhav Krishna, has built an AI recruitment platform that operates across 920 cities — serving precisely this overlooked workforce.
By abandoning the assumptions baked into conventional hiring technology — digital fluency, stable connectivity, standardized language — Vahan has engineered something rarer than a slick product: a system that actually works for Bharat. In this conversation, Krishna explains how they rebuilt the AI stack from the ground up, why language localization is the industry’s most underestimated problem, and what it truly takes to engineer trust with a worker whose livelihood depends on getting it right.
Founder and CEO
Vahan.ai
CIO&Leader: Vahan operates across 920+ cities with users who have low digital literacy and patchy connectivity — what did you have to fundamentally rebuild in your AI stack to make it work at that last mile?
Madhav Krishna: Most AI systems today are designed assuming users are digitally fluent, always online, and comfortable navigating apps. That assumption does not hold for a large part of India’s workforce.
So we had to rethink the stack around accessibility and inclusivity rather than sophistication alone. For us, AI is valuable only if it can remove barriers.
The first shift was moving to a voice-first architecture. Many of our users are far more comfortable speaking than typing, which is why our AI recruiter works over a simple phone call instead of requiring app-heavy workflows. A worker should be able to find a job, qualify for it, and complete onboarding without requiring high digital literacy or even a smartphone.
The second challenge was multilingual adaptability. India’s workforce doesn’t use standardized language patterns. Workers switch between dialects, Hindi-English mixes, regional slang, and hyperlocal expressions within the same conversation. We therefore fine-tune our systems on real-world recruitment interactions collected through our calling operations rather than relying on generic internet datasets.
And finally, we had to optimize for low-connectivity environments. Conversations often happen in fragmented sessions with unstable networks, interruptions, and dropped calls. So resilience and continuity became core engineering priorities.
CIO&Leader: Most AI hiring platforms are trained on white-collar, English-language data — how did you approach building training datasets that actually reflect the reality of a delivery worker in Tier 3 India?
Madhav Krishna: The biggest challenge is that most existing AI models are not trained on the realities of India’s frontline workforce. A delivery worker in Tier 3 India doesn’t communicate in polished resume language or standardized digital formats. Conversations are often multilingual, highly contextual, and shaped by regional dialects, literacy levels, and informal speech patterns.
While we don’t train models entirely from scratch, we do fine-tune existing ones using data collected through the calling operations conducted by our recruitment partners. This real-world interaction data is extremely valuable because it helps improve the models’ understanding of regional languages, local dialects, literacy patterns, and user intent — all of which are critical when engaging with gig workers who may not use standard digital interfaces.
We’re also actively experimenting with open-source models to build our own small language models (SLMs) tailored to our use case. We believe this approach holds enormous promise for India because these models are smaller, faster to deploy, and significantly more cost-efficient to run. That becomes extremely important when you’re trying to scale AI-powered services across a highly diverse and price-sensitive workforce at Bharat scale.
CIO&Leader: Going from a working prototype to 920+ city deployment — where did the model break first, and what was the fix that unlocked scale?
Madhav Krishna: The first thing that broke was not language understanding — it was workflow continuity.
In demos, conversational AI feels magical because interactions are linear and controlled. Real-world hiring is chaotic. Workers disconnect midway, switch languages mid-sentence, call back hours later, or hand the phone to someone else entirely. Our early systems struggled because they treated hiring as a single conversation rather than an ongoing worker journey.
The breakthrough came when we redesigned the orchestration layer around persistent context and human-in-the-loop workflows. AI now handles a large portion of repetitive recruitment operations, while humans step in where judgment, escalation, or intervention is needed.
That hybrid architecture unlocked scale by combining the efficiency of automation with the flexibility of human recruiters.
CIO&Leader: Infrastructure, language localization, last-mile access — which of these was the hardest engineering problem, and which one do most people underestimate?
Madhav Krishna: Language localization is by far the most underestimated challenge.
People often assume localization means translating a system into multiple Indian languages. But the real complexity lies in understanding how people actually speak across India.
Every language has multiple dialects, accents, pronunciations, and regional variations. Hindi spoken in Bihar sounds very different from Hindi spoken in Delhi or Rajasthan. The same applies to Tamil, Bengali, Marathi, Telugu, and virtually every major Indian language. Workers also frequently switch between languages mid-conversation, mix English with local vocabulary, and use highly contextual phrases that traditional AI systems struggle to interpret.
Training AI to understand these nuances is extremely challenging because the system has to recognize not just words but also intent, tone, pauses, and conversational behavior across very diverse speech patterns.
A major focus for us has therefore been making the AI feel more human and conversational rather than robotic. The interaction has to feel natural and intuitive, especially for workers who may be speaking to an AI system for the first time. That requires continuous fine-tuning using real-world calling data so the models can adapt to regional accents, local dialects, literacy levels, and natural speech flow.
Infrastructure problems can eventually be solved through optimization. But building AI that genuinely understands Bharat, and can communicate in a way that feels familiar, accessible, and trustworthy, is a much deeper challenge. That’s where the real engineering complexity lies.
CIO&Leader: You’ve driven sourcing cost reductions of roughly 40% — beyond cost, what’s the metric your clients are now demanding that you weren’t measuring two years ago?
Madhav Krishna: The questions were simple: how quickly can you deliver qualified candidates, and how reliably can hiring operations scale?
Two years ago, most clients were focused on sourcing efficiency. The questions were simple: how fast can you deliver candidates, and how much does it cost? That was the primary lens for evaluating performance.
Today, that has shifted quite meaningfully. The conversation is no longer just about getting candidates in the door. It is about what happens after the hire.
We are now being evaluated on activation and retention. Are people actually showing up for work? How quickly do they start earning? Do they stay beyond the first few weeks, or do they drop off early? And more importantly, how stable is that engagement over time once they are placed?
That has become the real measure of quality. It is no longer just about whether a position is filled, but whether that hire becomes a productive, sustained worker.
What we are seeing is a clear shift from hiring efficiency as the metric to workforce quality as the outcome. And that is where the market is steadily moving.
CIO&Leader: When you’re selling AI-powered hiring to a large enterprise, what’s the proof-of-concept moment that converts a skeptical CHRO into a committed partner?
Madhav Krishna: The turning point usually comes when enterprises realize that AI is not replacing recruiters but is dramatically amplifying recruiter productivity and operational scale.
Our AI recruiter has already improved recruiter productivity by up to 3x. That allows recruiters to manage significantly higher candidate volumes and improve fulfillment responsiveness. It also significantly increases the number of workers who can be reached simultaneously, which is critical in high-volume recruiting environments where speed and scale directly impact fulfillment.
For large enterprises hiring across multiple cities, recruitment is fundamentally an operational coordination problem. They are dealing with massive candidate volumes, fluctuating demand, high drop-off rates, and fragmented hiring workflows. Once they see that AI can automate repetitive tasks like initial outreach, candidate qualification, follow-ups, interview coordination, and support at scale, the value proposition becomes very tangible.
What also changes the conversation is how seamlessly AI integrates into existing workflows. Unlike traditional software systems, which often require organizations to change their behavior completely, AI can adapt to existing workflows.
That’s especially important for large organizations, because many enterprises want to become AI-first but struggle to move quickly due to their size and operational complexity. Startups like ours have an advantage there because we can move fast, iterate rapidly, and build highly tailored AI-native solutions around real operational pain points.
Ultimately, the proof point is not just automation. It’s when enterprises realize that AI can make hiring more predictable, scalable, responsive, and accessible without removing the human element from the process.
CIO&Leader: A delivery worker in a Tier 2 city has no reason to trust a chatbot with their livelihood — how did you engineer trust into the product?
Madhav Krishna: Trust in this segment comes from outcomes and simplicity, not branding.
Workers do not care whether AI powers something. They care whether it helps them get a legitimate job quickly, clearly, and reliably. That’s why we designed the system to feel conversational and familiar rather than robotic. The AI speaks in the worker’s preferred language, explains the process clearly, answers questions naturally, and guides them step by step through the recruitment process.
Voice was especially important because it lowers the intimidation barrier significantly compared to forms or apps. And over time, trust compounds organically. Once workers successfully secure jobs through the platform, referrals and repeat engagement begin to scale naturally within communities.
CIO&Leader: At what point does AI in blue-collar hiring stop being a tool and start displacing human judgment — and where do you draw that line?
Madhav Krishna: We see AI as an enabler that amplifies human capability, not as a replacement for people.
In blue-collar hiring, there are many repetitive, operationally intensive tasks, such as candidate outreach, qualification, follow-ups, interview coordination, onboarding support, and answering common queries. AI is extremely effective at handling these tasks at scale and with far greater efficiency.
But hiring is ultimately still a deeply human process. Workers often need reassurance, guidance, and trust, especially when their livelihoods are at stake. Employers also rely on human judgment in nuanced situations that require context and empathy.
That’s why we strongly believe in a human-in-the-loop approach, where AI handles 80 to 90 percent of repetitive operational tasks, while humans manage the remaining edge cases and nuanced interactions. The key is ensuring that this blended experience feels completely seamless to both the worker and the employer.
At Vahan.ai, our Vahan Leaders, who are recruitment partners on the ground, continue to play a critical role in building trust, guiding workers through the hiring journey, and managing local realities that require human understanding. AI helps make them significantly more productive and scalable, but it does not replace the importance of human connection in the process.
The goal is not to remove people from the hiring process. The goal is to use AI to make quality employment more accessible, efficient, and inclusive at scale.