hello! i'm partho das, a 19 y/o working on (and learning about!) AI systems and this website features some of my work.

more on my github and you can also find me on the internet by the username daspartho. also feel free to reach out at parthodas6176@gmail.com or @daspartho on discord.

i'm currently interning at anarchy (YC W23) and made open-source contributions to their LLM-VM project before that.

some fun projects i made along the way:

• prompt extend: tool to help with prompt engineering for stable diffusion models. also received grants from algovera.ai for it.
• MagicMix: implementing the MagicMix paper. also added a community pipeline to the diffuser's library.
• predict subreddit: classification model for predicting subreddit based on the title of a post.
• bored ape diffusion: diffusion model for unconditional image generation of bored apes.
• is it huggable: image classifier that determines whether something is huggable or not.

as for hackathon projects: built DistillClassifier for easily generating synthetic data for classification tasks using LLMs and dreamixer for creating comic strips from just text description.

also actively contributed to the huggingface transformers and diffusers library in the past.

did some mechanistic interpretability work investigating into circuits implemented by the gpt-2 small model and summarized my findings around localizing the model’s computation for the pronoun prediction task.

worked on exploring synthetic data interventions to reduce sycophancy in LLMs and reinforcement learning from language feedback in the SPAR research program.

also collaborated with AI safety @ UCLA's alignment research team on mechanistic anomaly detection project.