Naymul Islam

I work on whether language models stay safe when the words change. Safety training is done almost entirely in English. It does not always survive a change of language, register or framing, and almost nothing measures that outside English. I build the benchmarks that do, and the systems they run on.

Three papers so far, one as first author. The newest holds a harmful request fixed and changes only how it is written: 879 requests, five phrasings each, 18 models, 15,822 judged responses. Written as newspaper prose rather than as a casual message, the same request is answered 17.5 points more often. The test language is Bengali. The failure is not.

With BanglaLLM, where I am one of two research leads, we have released 224,000 Bangla instruction examples and an open model family from 1B to 13B parameters. Bangla has roughly 240 million native speakers and almost no evaluation of its own.

The other half is engineering. At Perspectivity I am a founding member and lead engineer. I built the ingestion and event-clustering pipeline, the orchestration around it and the Kubernetes deployment, and I run the production cluster. It has ingested more than 400,000 source articles and serves about 58,000 requests a day, with 186,000 web users over thirteen months. Live on the web, with iOS and Android in beta.

Before that: an agentic browser at Fellou AI, an autonomous coding agent at Menlo Park Lab, speech recognition and synthesis models at Freedoms.ai, and the data platform at AIMarketCap, where I lead the technical development. I write open source too: BrowserPilot, and merged work in Kubeflow Kale, Jina AI, browser-use and the CNCF Cloud Native Glossary.

None of the research is mine alone. I work with Abdullah Khan Zehady, Shubhashis Roy Dipta, Santu Karmaker and Nusrat Jahan Lia at BanglaLLM.

One small thing, since people assume: the ai in ai-naymul is not artificial intelligence. It is Chittagonian for I am. Every username I actually wanted was taken, and I like the dialect.