AI engineer building safer agents and imaginative products.

I’m Ronith Rashmikara, a Sri Lankan AI and software engineer working across agent safety, generative learning, fine-tuned language models, full-stack products, and immersive intelligent systems.

AI engineering projects and evidence

My recent work includes DataAgent-SafeBench, an open AI-agent safety benchmark with 120 controlled cases and 720 scored runs; Yume, a live generative language-learning world; and PersonaBench Aiko, a controlled character-model fine-tuning study.

I received the Best Presenter Award for the AI track at JASPER 2026 for presenting DataAgent-SafeBench.

  • 720 scored AI-agent safety runs (120 cases × 3 open-weight models × 2 repetitions)
  • 120 controlled benchmark cases
  • Fine-tuned language-model research with held-out evaluation
  • Live generative AI and immersive-web products

Start with the featured build, Project Alpha, then explore PersonaBench Aiko and KOI: Spirit Current.

Selected AI, software, and research work

Explore deployed products, research studies, live demos, public source code where available, technical decisions, and measurable outcomes.

  1. Yume

    A language-learning game on Visko Orbis: speak Japanese or English and the live world transforms, with quests, a sticker book, and a magic door.

  2. DataAgent-SafeBench

    A benchmark for authority confusion, prompt injection, and unsafe task escape in AI data agents, with a published v1.0.1 re-analysis.

  3. Orbit Personal OS

    A personal Android OS with local music streaming, audio-reactive artwork, daily planning, and private creation tools.

  4. Project Alpha

    A full-stack freelance marketplace with projects, bids, simulated escrow, reviews, disputes, and administration.

  5. PersonaBench Aiko

    A controlled LoRA fine-tuning experiment measuring character consistency against prompting baselines.

  6. KOI: Spirit Current

    A third-person browser experience inside an upside-down Gaussian-splat spirit garden.

  7. Ceramic Road Journey

    A rainy Gaussian-splat village adapted for browser delivery: 67 MB of splats down to a 2.47 MB default tier.

  8. JASPER Textile Lab

    A computational study of Japanese and Sri Lankan textile design DNA.

AI and full-stack engineer

I work across AI agents, machine learning, large language models, generative media, full-stack engineering, world models, and immersive web technology.

Contact Ronith Rashmikara or review the source code on GitHub.