The model is the easy part.
You call an API, you get a response. The capability is mostly there. The hard part — the part that actually decides whether something is useful — is everything surrounding it.
How do you explain to someone what the tool does when they've never seen anything like it before? How do you handle the case where the model is confidently wrong? How do you build something that works on a ₹5,000 phone with a 3G connection in a moving auto-rickshaw?
These are product and engineering problems, not model problems.
Most AI tools I've seen are built by people who are excited about the technology. That's fine. But there's a ceiling to how useful something can be if the primary axis of optimization is "impressive demo." Impressive demos are for engineers. Products are for people.
The people I want to build for don't think about language models. They think about their problem: filling a form, understanding a document, getting an answer to a question they're embarrassed to ask a person. The technology should disappear entirely.
This is where I spend most of my time: the unglamorous middle layer. Error messages that actually explain what went wrong. Fallbacks that don't break the experience. Latency that doesn't feel like waiting.
None of it is glamorous. None of it makes a good demo. All of it is the difference between something people use once and something they come back to.