We build the unglamorous half of AI

We started in 2019 building search and data systems, well before language models made retrieval fashionable. When the current wave arrived, we recognised the problem immediately: the hard part was never generating fluent text. It was making sure the right information reached the model, and knowing when it had not.

That bias shapes everything we do. We build evaluation harnesses before we build features. We measure retrieval quality separately from answer quality, because conflating them hides the actual fault. And we design systems that refuse to answer rather than approximate, because in telecom billing or clinical documentation an invented answer is worse than no answer.

We are roughly thirty people across engineering, research, design and delivery, headquartered in Dhaka 1219 and working remote-first with clients across seven industries. Most engagements that start as a single project become long-term relationships — our client retention sits at 94%, which is the number we are proudest of.

We are not the right partner for everyone. If you want a demo by Friday, there are faster options. If you want a system that still behaves correctly in eighteen months when your documentation has changed and your traffic has tripled, that is the work we do.