Manav

Our paper on localizing ASR hallucinations has been accepted to EMNLP 2026 Main conference! This is my first main conference paper. Big shoutout to my co-authors and see y'all in budapest 🫡

Production calls are different from demo calls, and it changes everything. @kogtafinancial found that difference with Smallest AI. They'd tried other partners before, but the agents never sounded human enough for real calls. With Smallest, they found voices that sounded human, low enough latency for conversations to flow naturally, and agents that remembered the customer's details and could respond on the fly. What started as testing quickly became production-ready, with agents handling real borrower conversations and collecting payments on the call itself. Here's @kogtafinancial's CTO on the journey.

How team culture gets preserved at @smallest_AI ft. a screenshot from our internal WA group. One thing I’m constantly in awe of is how openly our founders show appreciation for the team. There’s also a real humility across the system. The culture has always felt like: we’re tackling this together, and if you need help, I’ve got you. Smallest has grown from essentially 0 → 70 people in two years, but the unfiltered stories from day one are still being told and we continue to dedicate an hour to really get to know every person who joins (if ykyk). What an incredible team to be building with.

Hosted a boulder & bond session. We met on twitter - edition bouldering. Fun muchos 🤷‍♂️ @dawninthewild @pahwaginni @geetikaguptaa @KastAmeya @18RawatG @kunalvats_ @sanjkrao @manavslab @vaish0709 @huesofbanter @_namita_r @ayush1599

Kogta Financial, one of India's fastest growing retail-focused NBFCs, serves 100,000+ borrowers across 300+ branches pan India. Debt collection in India is incredibly diverse. Borrowers switch between languages, dialects, and local ways of speaking. Even a single mispronounced number can change the outcome of a call. That's exactly the challenge they set out to solve. @kogtafinancial built their collections agents entirely on Smallest AI's voice stack, giving them native support for rural accents, mixed-language speech, and numbers spoken in regional languages. The impact: - 15M+ calls placed, with high value EMIs settled during the call itself. - 130+ concurrent calls at a 0.2% failure rate.