Sajal Regmi
Applied AI engineer
I like working with large language models — both training them and building products on top of them.
I trained Arkios-1B, a 1.04B-parameter bilingual English–Nepali model, from scratch in a C/CUDA stack I wrote. On the application side I build Arkios, an enterprise platform for deploying AI agents.
Research
Arkios: An Open Bilingual English–Nepali Language Model Trained From
Scratch, with a Devanagari-Aware Tokenizer
arXiv preprint, 2026
Vowel Signs Are Not Letters: A Pre-tokenization Ceiling on Multilingual
Tokenizer Fertility
arXiv preprint, 2026
GPT Semantic Cache: Reducing LLM Costs and Latency via Semantic
Embedding Caching
arXiv preprint, 2024
Open source
Arkios Tokenizer
65,536-entry English–Nepali BPE tokenizer for generative models, with a
harness that regenerates every published number on a laptop CPU.
GPT Semantic Cache
Caches LLM responses by semantic embedding similarity to cut latency
and inference cost.
Transformer from Scratch
Ground-up PyTorch implementation of the architecture behind modern LLMs.