hey, i am arjan.

16yo . security researcher . founder .

i build and break things for a living. at 12 i was building nepfolk, which i never launched. at 13 i co-founded glow tech and scaled it to 10+ clients a month. at 14 i built and exited ggamechamps and hacked google, ea, stanford, shopify, jetstar, nba, msi & more. right now i'm co-founder & cto at heedeo (ef f26), the trust layer for the agent economy. we prevent free trial abuse. also running arniko hack club (400+ members) in nepal.

what i've been up to:

• co-founder & cto at heedeo (ef f26). trust layer for the agent economy. we prevent free trial abuse
• built the #1 nepali tts & stt models at naamche labs. tts: swar (launch video). stt: kriti (launch)
• worked as vciso at dench (yc s24), helped them get soc2
• founded arniko hack club, a nonprofit with 400+ members. through it, organized one of asia's largest teen hackathons (daydream biratnagar) and nepal's largest teen game jam (campfire biratnagar), where 120+ students shipped games overnight
• ran invention novelty, a 30-day founder residency in bengaluru, bringing ambitious founders and researchers together with housing, workspace, compute, and technical mentorship
• built and exited ggamechamps, a gaming clips + ml datasets platform. got a $5k grant from emergent ventures for it
• found 100+ vulnerabilities across hackerone, bugcrowd & private programs. hacked google, ea, stanford, shopify & more
• #2 in jetstar hall of fame
• one of the youngest to be assigned a cve-2025-51588
• scaled glow tech to 10+ clients a month at 13

things i've worked on:

dench arniko hack club daydream biratnagar invention novelty ggamechamps shipsecai glow tech cyber alert nepal nepfolk

side projects:

naamche labs — ai lab building for the nepali internet swar · #1 open nepali text-to-speech. clone any voice from a short voice note kriti v1 — open nepali speech recognition model, ranked #1 kriti telephony — asr model trained on telephony-grade nepali audio pagepatch — permanently edit any website. changes survive refresh.

papers:

merge-based vocabulary extension stalls at the pre-tokenizer: an english-preserving repair for nepali, tested in continued pretrainingai models like llama and qwen read nepali in up to 5.8x more pieces than english, because a hidden rule splits words at every vowel sign. a small fix to that rule makes nepali as cheap as english, without changing how the model reads english.solo author · arxiv 2026 · code · tokenizers for qwen3 & llama 3.2

baselines before architecture: evaluating coding agents for autonomous penetration testingcan off-the-shelf coding agents like codex hack apps on their own? on 104 hacking challenges they already solve a large share, sometimes matching custom-built hacking systems, so new systems should be compared against them first.with ananda dhakal & krish neupane · arxiv 2026 · 60k+ views on x