How Nclude.ai turned broken portals into completed applications

Nclude.ai didn’t set out to fix the internet’s code; it set out to make the internet usable. Built by Incskill / I-STEM co-founders Nishit Bhasin and Kartik Sawhney, the system converts hostile PDFs, fixes broken forms, and delivers the same service over web, WhatsApp, and IVR. In India, it now rides a government channel, Sugamya Bharat, so a benefit application can be filed by phone when the site won’t load.

Most accessibility efforts still chase compliance at the source. Teams file tickets, tweak markup, hope a screen reader cooperates, then repeat on the next site. Right intention, wrong unit of progress. Bhasin’s contribution is a user-first, above-the-stack layer. Nclude converts hostile content, completes broken workflows, and delivers the same service over web, WhatsApp, and plain phone lines. The result is less “please fix your code” and more “the textbook is readable, the form is submitted, the benefit is applied for.”

The origin story is pragmatic. India’s disability community runs into the same walls daily. Multi-column PDFs shred under OCR. STEM diagrams arrive without alt text. Job portals trap keyboard users in two fields. Bandwidth is fragile. Devices are shared at home. Bhasin’s product brief read like an operations plan. Build a conversion pipeline that handles messy government and education documents at scale. Add an agent that fuses computer vision with DOM understanding, maps intent to form semantics, and executes steps with the user in control. Ship it where people actually are. Web for power users, WhatsApp for low friction, IVR for anyone with a basic phone.

What separates this from the NGO project archetype is how Bhasin links products to real-world gatekeepers. Microsoft’s AI for Accessibility award underwrote the remediation portal and document AI backbone, which forced serious testing discipline for complex PDFs and STEM content. The Turn.io and OpenAI accelerator put WhatsApp live on an eight-week clock with monthly active user gates, so no open-ended pilots. A Washington State Department of Services for the Blind contract moved the work into operations with statements of work, reporting, and fee schedules. UNICEF and UNDP visibility pushed the model outside India into low-bandwidth contexts where IVR is not a fallback but the primary interface. Pilots with Bosch and Meta explored two edges of the same problem: career discovery grounded in disability constraints and task automation on enterprise HR portals.

Bhasin’s most important choices are product choices. Treat disability constraints as first-class features, not footnotes. The career module does not give motivational speeches. It evaluates assistive tech compatibility by role family, remote feasibility by city, commute realities, accommodation effort, task shape, and the ramp to first income. It ranks roles the user can plausibly do, then links directly into assisted submission, so the plan is executable today. That sounds obvious until you try to ship it in the wild. Shared devices require privacy that actually works, so the team added PIN-locked sessions and local redaction. Foundation models carry ableist bias, so they built prompt policies and classifiers that block pity narratives and require source citations. The agent exposes every step with confidence indicators, a why behind each action, an undo, and escalation to a trained human when the site’s hostility crosses the line.

Government integrations are where ideas earn or lose their stripes. The Sugamya Bharat channel matters because it flips the unit of distribution. Instead of persuading every ministry, state board, and registrar to remediate their own interfaces perfectly, content routes through one national access layer and publishes parity over the web, WhatsApp, and IVR. That is the difference between pamphlets about accessibility and infrastructure that quietly reduce failure rates. The Washington State engagement proves the stack can survive procurement, audit, and day-to-day delivery in a very different regulatory setting. This is not vanity. If disability tech is more than a demo, it must live where invoices get paid and programs renew.

The obvious critique is that none of this fixes the root cause. Websites still ship broken. True, and that is the point. Bhasin’s model does not replace standards. It gives the ecosystem a pressure valve that moves outcomes while the rest of the web catches up. When the agent finishes a benefit application by voice for someone without a smartphone, the moral accounting changes. When a blind student studies a properly structured chapter the night before an exam because the PDF was repaired upstream by machine and audited downstream by a human, the standard conversation gets leverage it never had.

There is a second-order effect that is easy to miss. An above-the-stack layer generates a corpus no one else has. The documents governments actually publish. The form patterns HR portals actually use. The failure modes that kill submissions. The paths that succeed. The languages and dialects that drive IVR completions. That data becomes a moat. Competitors can copy UI. They cannot quickly reproduce the evaluation harness for ableism and hallucinations, the task trees and gold paths, the privacy playbooks for shared devices, or the acceptance criteria regulators will sign. This is how a social-impact startup stops being a grant project and starts operating like a product company with defensible assets.

The field is already shifting. Disability desks and nonprofit partners are rewriting KPIs from counseling volume to submissions, interviews, and income. University teams are using the same stack for exams, scholarships, and aid workflows. International pilots are not exhibitions. They are translations of the same playbook into places where bandwidth is thin and phone lines carry most of the load. None of this changes the fact that the work is unfinished. It does show a credible template for institutions that need outcomes now and cannot wait for the public web to behave.

Strip away the brand names and you get execution with scope. Bhasin, as Head of Product, owned the architecture and the shipping cadence, tied capital to modules, and kept analytics honest. Kartik Sawhney, a respected accessibility leader, co-drove partnerships and policy alignment so deployments could land where they matter. Together, they treated safety, privacy, and dignity as product requirements, not aspirations. They aimed for integrations where millions live, not where a few might click.

There is plenty of noise in AI for social impact. This reads like a signal. A platform that is most useful when screens, networks, and policies failing is easy to underestimate. It is also the kind of system that, once deployed, becomes invisible and indispensable at the same time. If you want a test for whether an accessibility product actually matters, use this one. When the network is bad, the site is broken, and the stakes are high, does the user get through? In Nclude’s world, built by Incskill Inc aka I-STEM, the answer is yes.