A blind technologist’s journey to building Samora AI, with OpenAI in the stack

Kartik Sawhney's relationship with technology did not begin in a Stanford lab or a startup accelerator. It began in a school corridor in Delhi, when a rulebook told him he wasn't allowed to study science because he was blind. At the time, the Central Board of Secondary Education did not permit visually impaired students to take the science stream. Sawhney appealed directly to the board chairman, and won, opening the door for blind students across the country to study science. That early victory shaped how he approached every barrier thereafter: when systems aren't built for you, challenge the system, not your ambition.

Accessibility as infrastructure

Winning access, however, did not erase the structural gaps that followed. None of the textbooks were accessible, so Sawhney typed them out manually, four hours a day. Later, when his IIT applications ran into similar restrictions, he chose to leave India to study computer science at Stanford. Each step reinforced a belief that would stay with him: accessibility shouldn't depend on individual workarounds; the tools themselves should work for everyone.

That conviction shaped his first venture, I-Stem, which digitised inaccessible content, provided remediation services, and worked closely with students and institutions, including CBSE, to make learning material usable for people with disabilities.

The work gave Sawhney an unusually sharp view of where mainstream systems fail, especially in voice. As someone who navigates most interfaces through a screen reader running at 20x speed, he depends on voice in ways many users do not. Commercial voice assistants, however, remained stuck at a surface level. They misheard accents, struggled in noisy environments, and failed to hold context beyond a sentence or two.

From assistive tech to enterprise voice

"If we've had voice AI for years," Sawhney asks, "why are we still only using it for weather updates?"

The question emerged from a series of practical problems his team encountered while building assistive technologies: phone lines that dropped context mid-conversation, voice agents that couldn't handle Indian accents, models that broke as soon as background noise entered a call, and datasets that failed to reflect the everyday complexity of Indian speech.

"We realised voice is powerful," he explained. "But making it work in India, with our telephony networks, dialects, and call patterns, meant we had to build custom tooling at every layer."

That realization eventually led to Samora AI, the company Sawhney now heads as Co-founder and CEO.

The gap Samora set out to solve was far larger than improving consumer assistants. India's service economy runs on conversations: education counsellors calling students, real estate developers following up with leads, mobility operators managing fleets, NGOs running outreach campaigns.

Yet call centres struggle with scale, quality, and cost. A team of 20 can place only so many calls. No human team can reach half a million people in a weekend. Lead qualification decays when messages don't go out on time. Meanwhile, users speak in mixed languages, over traffic noise, often interrupting mid-sentence. Most voice bots fail almost instantly under that pressure.

Engineering for chaos

Samora was built to withstand it. Its design begins with how people actually speak. Sawhney’s team studies pauses, fillers, hesitation, emphasis, and the rhythm of everyday dialogue to make voice agents sound less mechanical and more conversational. But naturalness means little without robustness, and real-world robustness is where most systems break.

India's telephony networks are inconsistent, acoustic environments are unpredictable, and language use varies dramatically from region to region. Samora's engineering layers, which include speech recognition, reasoning, telephony logic, analytics, and human escalation, are designed to absorb this variability instead of collapsing under it.

The platform relies on OpenAI's Whisper and real-time speech models to handle transcription across accents and background noise with reliability the team could not achieve earlier. GPT-4 and GPT-5 power the reasoning layer, enabling agents to understand intent, retain context across conversation turns, and adapt responses based on how callers speak. OpenAI's expressive text-to-speech models provide natural pacing and tone rather than flat, robotic audio.

Because enterprises need safety nets, Samora includes warm-transfer capabilities, allowing human agents to step into live calls smoothly when required. OpenAI also accelerates Samora's internal development, from analyzing call logs to testing conversational flows, allowing the team to iterate faster and focus effort where it matters most.

From inclusion-first roots to large-scale deployments

Although Samora's origins lie in disability inclusion, the platform now extends far beyond that ecosystem. A disability app built on top of Samora continues to support more than 30,000 active users, bringing the accessibility mission full circle. But the technology has proven equally effective in commercial contexts.

One of its early deployments helped a university engage over 50,000 prospective students within five days, a campaign that also became a live testbed for refining messaging through real-time conversational A/B testing. Real estate companies have used Samora to revive dormant leads and accelerate qualification funnels. A UN agency has worked with the platform to voice-enable its youth employment services, expanding reach in ways traditional outreach couldn't match.

Across these use cases, organizations report that repetitive call workloads have reduced sharply, lead-handling cycles have shortened, and insights that once took weeks to assemble now surface within hours. "AI is reducing noise," Sawhney observed. "It's helping teams figure out quickly where to focus instead of spending days sorting through data."

What comes next for voice as an interface

Samora's recent admission into Y Combinator reflects both its technical depth and the scale of the opportunity ahead. The team is expanding its infrastructure, preparing for global deployments, and building more sophisticated conversational layers as OpenAI continues to advance its capabilities.

Sawhney sees a future where voice becomes the default interface across software, not only for customer-facing calls, but inside apps, workflows, and internal systems. "We're only scratching the surface of what voice AI will look like," he said. "Soon this won't just be about calls; it'll be how people interact with all digital systems."

For a founder who spent his childhood rewriting inaccessible textbooks, building AI systems that listen, understand, and adapt feels like a natural evolution. The tools that once required four hours of manual typing each day are now being rebuilt from the ground up; not as workarounds, but as systems designed for everyone. And with the pace at which voice AI is advancing, that vision is arriving faster than anyone expected.