we went live.
helios is live at Khalti.
i have been trying to find the right way to write about this for a few days and i keep landing in the same place. there is no framing that quite captures what it feels like to have something you have been building for months actually running in production, at scale, with real users depending on it.
so i will just write it plainly.
what live actually means
in this industry, "launch" can mean many things. a demo. a beta. a soft rollout to a small cohort. a press release that precedes the actual product by several months. we have done some of those things.
this is different. helios is running inside Khalti's production infrastructure. real KYC verification, happening in real time, for real users trying to complete real financial transactions. not a pilot. not a proof of concept. production.
the number of times i have said the word production in the last week, as confirmation to myself that it is real, is higher than i am comfortable admitting.
how we got here
the months between when we started building helios and today do not compress easily into a summary.
circa march 2026
there was the core research. building a system that could verify identity in a way that actually worked for this market, with the document types, the script variations, the image quality distributions, the edge cases that generic global providers were not designed to handle.
there were the deployment decisions. helios runs on-premises. it had to. the data residency requirements were non-negotiable from the first conversation, and the architectural choices that followed from that constraint shaped almost every subsequent decision. you cannot shortcut that. you either design for it from the beginning or you spend a very painful six months redesigning later.
there were the integration cycles. connecting helios to existing infrastructure is not a simple task. the banking and fintech systems in this market were built by different teams over different timelines with different assumptions. making helios fit into that without disrupting what already works is most of what the integration work actually looks like.
there were the QA cycles. the evaluation sets. the threshold calibrations. the false positive rates and false negative rates and the conversations about where to draw lines when there is no objectively correct answer, only tradeoffs.
and there were the conversations. with the teams at Khalti, with compliance, with the engineers, with the people who would ultimately be responsible for what happened when helios made a decision that someone had to stand behind.
all of that compressed into today.
what i actually felt
i did not feel what i expected to feel.
i expected something closer to relief. we built the thing, it works, we shipped it. relief is the natural emotion.
what i actually felt was something quieter and more serious. closer to responsibility than celebration.
when helios is running in production, it is making decisions that affect people's access to financial services. a user trying to verify their identity to complete a transaction is depending on the system to work correctly. if it does not, the failure is not abstract. it has a consequence for a specific person at a specific moment.
i think about that a lot. i thought about it during the build and i think about it more now that the system is live. it sharpens how i think about every design decision we make, and it sharpens how i think about the work that still needs to happen.
going live is not the end of anything. it is a change in the nature of the obligation.
what is next
helios does not stop developing because it is in production. in some ways the most important part of the work starts now.
the evaluation sets we built during development are good but they are not complete. no evaluation set is complete. production will surface edge cases and failure modes that even the most rigorous internal testing did not anticipate, and those discoveries will drive the next phase of the research and the next set of improvements.
the reliability work continues. the monitoring infrastructure has to catch things before users do. the feedback loops between what we see in production and what we improve in development have to be fast and well-designed.
and the deployment work expands. Khalti is one institution. the problem helios is solving is not specific to Khalti. it is the same problem that exists across the fintech and banking ecosystem in this market, and the solution that works here is not fundamentally different from the solution that would work elsewhere.
we are already having those conversations.
a note on the team
ujjwal and firoj have put an unreasonable amount of themselves into helios.
there is a version of this essay that is more personal about what the last several months have looked like inside the team. the late nights, the disagreements, the rebuilds, the moments where things were not working and it was not clear when they would. i am going to keep most of that between us. but i want to say clearly that whatever helios becomes, it is the product of three people who were genuinely committed to getting it right, and the fact that it is live and working is as much about them as it is about anything else.
we went live.