Built by one person, for creators who care about their viewers.
Virtual turns screen recordings into interactive knowledge bases, viewers ask questions at any timestamp, confusion gets mapped, and creators answer once for every future viewer who asks the same thing.
Who built this
Virtual is built and operated solo by Saarthi Marwaha, based in India. Every line of code, every design choice, every customer email, it's all me. The advantage of a solo founder: you talk to the person who can actually fix what's broken. The disadvantage: it's one person, so reply times honour the time zone of one human.
I started Virtual after watching creators repeatedly answer the same five questions under every video they posted. Comments sections aren't searchable. Course platforms bury Q&A. Discord channels lose context after a week. A video viewer with a question at 3:47 has no good place to put it, they either give up or DM the creator and add to the unread pile.
Virtual is the answer I wanted to build: a layer that sits on top of any screen recording, listens for viewer questions at specific timestamps, answers what it can from the transcript + frame analysis + the creator's saved ground-truth replies, and quietly clusters anything that keeps coming up so the creator only has to answer it once.
How the product is made
The stack, in case it matters to you:
- Frontend: React + TypeScript + Vite, hosted on Vercel.
- Backend: Fastify (Node 22) running on Railway, with a BullMQ pipeline for video processing.
- Database: Postgres on Supabase, with pgvector for semantic search over transcript chunks and frame descriptions.
- Storage: Cloudflare R2 for video files and extracted keyframes.
- AI: OpenAI Whisper for transcription, GPT-4o-mini for frame analysis, Anthropic Claude Sonnet for Q&A answers, Groq Whisper as a low-latency fallback. We semantically cache answers so repeat questions return without re-calling the LLM.
- Email: Resend for all transactional mail, sending from the verified
heyvirtual.medomain.
Every viewer question runs through a multi-stage retrieval pipeline before any answer is generated. The point is to never make up a frame or claim something happened in the recording that didn't. If the evidence is weak, the answer says so out loud.
Where AI is used (and where it isn't)
Per Google's guidance on AI-generated content, here's the honest breakdown so you don't have to guess:
In the product
- Viewer questions are answered by Claude Sonnet using the recording's transcript and frame analysis as context.
- Keyframes are described by GPT-4o-mini. Descriptions are cached so repeat use is free.
- Transcripts are produced by OpenAI Whisper (or Groq Whisper as a faster fallback).
- Question similarity (for the "Master Doubts" clustering and the semantic cache) uses OpenAI's text-embedding model.
- Creators can write their own ground-truth answers, those are used verbatim instead of any AI output when a viewer asks something similar.
On this marketing site
- Page copy (landing, pricing, About, policies) is written by me, the founder, and edited by me.
- Some longer-form blog and comparison pages use AI as a drafting and structure tool, then are reviewed, fact-checked, and edited by hand before publishing. Anywhere this happened, the page carries my byline.
- No content is published unread or unreviewed.
- Product screenshots and diagrams are real captures from the live product, never AI-generated images of a fake UI.
How we handle your data
Short version: your recordings are private by default, shared only by per-recording links you control, and we never train any model on your private content. Full details on the Privacy Policy page.
Get in touch
Email team@heyvirtual.me for anything, bugs, feature requests, billing, partnerships, or a quick hello. It lands in my own inbox; I read every one. If you're signed in, you can also use the in-app support form.
