How FOSS and AI are Changing Government Workflows at NHAI

In 2025, IndiaFOSS was the first in-person developer event I ever attended. Right after returning to Delhi, I joined the AI Cell at NHAI, the National Highways Authority of India. So it is kind of a full circle moment as a year later, after a lot of trial and error, this talk is supposed to be a reflection on what we have been able to build, what surprised us, and what FOSS made possible inside a public-sector organization.

Government organizations are often assumed to be slow adopters of new technology, and for good reason, their structure has those constraints and inertia to make them so. At the same time, outside of coding tools, enterprise AI adoption for knowledge work products has been slow, shallow, or stuck at the demo stage. 

Our experience at NHAI has complicated both assumptions.

With a small team of designers and engineers, working closely with NHAI insiders, we have been able to build a set of internal AI and productivity tools for the organization. 

Marg Sarthi, NHAI’s internal AI assistant, is the flagship example. In production, it has crossed 15,000+ conversations, 50,000+ questions being asked by 1,000+ internal NHAI employees, with 2 out of 3 being repeat users coming back at least once a week.

That gave us a chance to ask why this worked as a daily workflow tool when many internal AI assistants struggle to gain their footing beyond demos. People use it not as a generic chat interface, but for work they already do: finding policy circulars, interpreting clauses, drafting notes and letters, querying project data, working with scanned data heavy PDFs, and navigating approval-heavy processes.

Along with this, across NH-AI, we have built smaller tools as well, from document utilities to workflow-specific assistants. A pattern that we see emerging: each useful tool had builders on one side and an organizational counterpart on the other who pushed for these tools and iterated with us quickly. These institutional co-creators helped translate messy internal needs into product direction: what users would trust, what workflows mattered, and where governance boundaries could not be hand-waved away.

FOSS has mattered tremendously in us being able to make this possible, because it gave this small team agency, often rare in a process led traditional organization. In a closed vendor-led setup, every meaningful change can become a request: wait for procurement, wait for roadmap alignment, wait for customization, wait for integration support. In our case, the ability to run, inspect, adapt, and extend open systems meant that user feedback could turn into working software quickly. 

One particular OSS - LibreChat became the base of a tool like Margsarthi and gave us a hackable foundation for the assistant along with MCP giving us a way to connect tools and data without rebuilding the whole chat layer. Forgejo and self-hostable observability tools helped us keep ownership of the development and feedback loop.

This led to us working with proven primitives and being able to ship faster. AI has made this even more powerful. A small team could understand unfamiliar open-source codebases faster, customize them around our workflows, and build tools that would have been too slow to attempt if every change depended on an external vendor cycle.

We do not think that we have a final playbook for government AI, but we have arrived at hints of it as we do have a real production story from inside one public institution: how FOSS and AI helped us move faster, customize deeper, avoid vendor-roadmap bottlenecks, and build tools close enough to daily work that people actually used them.

This talk shares what we learned from that journey, and why the future of AI in government may depend as much on open infrastructure and institutional co-creators as it does on base models.

• One lesson from this work is that adoption is often a timing problem. In large organizations, the need may be real, but by the time a vendor cycle, procurement loop, or customization request produces something usable, the workflow has shifted or users have already found a workaround.

• We want attendees to see how FOSS changes that loop: not just by reducing cost, but by giving institutions the ability to inspect, adapt, host, observe, and replace the systems they depend on. Because the team could build on open, extensible systems and stay close to NHAI stakeholders, feedback could turn into working software quickly enough for users to try it in their daily workflows.

• The talk will also cover why internal co-creators matter. NHAI insiders helped translate organizational needs, trust concerns, and governance boundaries into product direction, so the tools were not just technically possible but institutionally usable.

• Finally, we want to make the case for internal AI capacity in government. Vendors can be useful, but transformation cannot be fully outsourced. When a small internal team of designers and engineers works with institutional co-creators and owns enough of the stack, it can adapt faster, learn from users directly, and build systems that are accountable to the organization rather than only delivered to it.