Hi, I'm
Vipul Srivastava
Engineering | PolicyMe
I like building complete products end-to-end — from the first idea to something you can actually use. Lately that's meant an AI news service, a loan-cost tool, and a handful of LLM playgrounds that run entirely in your browser.
I build things I'd want to use myself, and I ship with AI coding tools in the loop the whole way. The playgrounds below are a small taste of that — pick one and it runs on your own GPU, no server, no API keys.
Playgrounds
Four things you can run right here.
Each one runs on your own GPU via WebGPU — no server, no API keys. Click a card to open it.
🖥️ In-browser coding agent →
A Claude Code–style agent that edits a virtual project in a terminal, running a tiny open-source model live on your GPU. Ask it to change a file and watch the diff.
🔍 RAG playground →
A from-scratch RAG pipeline with every step drawn on screen: chunking, embeddings, cosine retrieval, a live token-budget compaction bar, and a real tool loop.
🧪 Distillation playground →
A teacher and a student model answer the same prompt side by side, with latency, token counts, a teacher-as-judge chart, and a one-click dataset export.
📊 Evals playground →
What an LLM eval actually is: exact match, fuzzy token-F1, and LLM-as-a-judge, run over a real suite — with a confusion matrix and charts you can read.
🎛️ Modalities →
Speech-to-text (Whisper), zero-shot image understanding (CLIP), and semantic search — running on your CPU via transformers.js. The one set of demos that needs no GPU.
Projects
Things I've built.
📰 NewsFlow ↗
AI-narrated news podcasts you can hear in 5 minutes. It turns the day's top headlines into a personalized audio briefing with synced subtitles — pick your country, categories, and language, then listen on the site or get it delivered by email. Free for everyone, no account needed.
🏦 LoanReady ↗
Know exactly what your loan will cost you. LoanReady takes opaque loan terms — interest rates, fees, and repayment schedules — and turns them into a clear, upfront picture of what you'll actually pay, so you can compare offers before you commit.
💬 vllm-chatgpt Live demo
A naive GPT implementation, built to understand all the fundamentals. A full-stack chat app (FastAPI + React) with RAG over uploaded PDFs, web browsing, MCP tool integration, and token-aware sliding-window context — served by Ollama or vLLM over PostgreSQL + pgvector. I ported the core to the RAG playground above.
🧪 vllm-model-distillation Live demo
An app for model distillation, built to understand what it's about. A playground that runs prompts through a teacher and a student model side-by-side via vLLM, stores every prompt–response pair, and exports them as structured datasets for training smaller, cheaper student models. The in-browser port runs both models on WebGPU.