> abhay kashyap_

An early experiment in fully voice-driven interfaces -- you talk to the UI and it reshapes itself in real time, no typing. I built it in early 2025 with Pipecat, when the tooling was still rough.

It began as a voice-browsed shopping experience on mobile, then moved to the living room -- voice as a TV interface that turns the remote into a pointer for the big screen. I showed some of it at AI Tinkerers.

It started as whisper-ui -- a simple Streamlit app to download and transcribe YouTube videos, built right as Whisper landed.

Later, I extended it into a toolkit to embed, index & search across the transcripts -- a knowledge layer for RAG. It picked up 700+ stars before I moved on.

Plumage was an early proof-of-concept for generating fashion designs, from apparel to body art.

It paired object detection with Stable Diffusion's inpainting model, and shipped as a platform to create, share, and try-on.

Me and my buddies, Xi & Reza, hacked this up in a couple of weekends with OpenAI's CLIP, Pinecone & Django.

We hoped for Lambos to magically show up after launch but they didn't so until then -- Ape. Together. Strong. 🦍 🦍 🦍

WildePod is an AI-powered, human-in-the-loop system to manage camera trap images and annotations -- built for Felidae Conservation Fund, a wildlife conservation & research organization.

I built this back in Fall '21 using Django, MegaDetector, Annotorious -- it was also my first major AI-assisted coding project (GitHub Copilot beta).

Rapalytics let you explore and compare lyrical features across rap artists -- built off 10k+ songs from 150+ artists with D3 & Django.

It accompanied my PhD research at the time & I haven't maintained it since, so it's dated.

Semantic Textual Similarity (STS) measures how close two sentences are in meaning. A system scores each sentence pair, and those scores are graded against human ratings. For much of the 2010s it was a key benchmark for sentence embeddings & large language models. It ran as a SemEval / *SEM shared task, where our lab competed in 2013 and 2014.

Our 2013 system used LSA-based word embeddings fused with WordNet and ranked best among 89 submissions. In 2014, we extended it to handle named entities via knowledge graphs, multiple languages, and text of different lengths, ranking first in several of the subtasks.

This became a 2015 journal paper and a core pillar of my PhD dissertation, where I tackled cold-start & long-tail for music recommendation using deep neural networks.