Skip to content
← All work

Short-form video · CreatorsWith Kenil JoshiLive

ReelForge

Raw footage in, captioned short-form video out — on its way to an editor that edits for you.

The problem

Upload a talking-head clip; ReelForge transcribes it, turns the speech into styled, animated captions, lets you drop B-roll and sound over any line, and renders a vertical cut ready for Reels and Shorts.

Short-form video rewards people who post every day and punishes the hours it takes: transcribing, timing captions word by word, hunting for B-roll, exporting in the right shape.

Most tools solve it by uploading everything to someone else’s render farm. That’s slow, expensive to run, and your footage leaves your machine. We wanted the fast part now — and a path to an editor that runs entirely in your browser.

FFmpeg pass for captions, B-roll, audio and watermark
1
FFmpeg pass for captions, B-roll, audio and watermark
experiments toward a fully in-browser editor
39
experiments toward a fully in-browser editor
priced by location, settled in paise
₹ / $
priced by location, settled in paise

How it works

  1. 1

    Upload

    Clips up to a minute. Phone rotation metadata is respected, a thumbnail is grabbed and the audio extracted to 16 kHz mono.

  2. 2

    Transcribe

    Speech-to-text with word-level timestamps. If there’s no speech, you’re told so — and the upload doesn’t count against your plan.

  3. 3

    Caption

    Words are grouped three at a time into caption segments you can restyle and animate — fade, slide, zoom, bounce — previewed live on a canvas.

  4. 4

    Layer

    Search stock video, images and sound, then pin each asset to the caption segments it should cover, so it lines up with speech by construction.

  5. 5

    Render

    A worker composes everything — B-roll, audio mix, captions — into a single FFmpeg pass and emails you when the video is ready.

Decisions I’d make again

Three words per caption

The caption segment is the unit of editing. Captions read punchy, and anything you attach to a segment is timed to speech automatically.

One filter graph

Overlays, audio mix, watermark and burned-in captions happen in one encode — no chain of intermediate files.

A free tier that pays for itself

Free projects keep B-roll in your browser and expire after seven days; paid features are enforced on the server, not trusted from the client.

Experiment before you build

39 standalone experiments de-risk the in-browser editor — WebCodecs, WebGPU, OPFS, on-device speech — before any of it touches the product.

The journey so far

From the first annoyance to telling people about it. Dates come from the commit history.

  1. Get annoyed

    Captions shouldn’t take an hour

    Every creator we knew spent longer timing subtitles than recording. That was the whole brief.

  2. 4 Oct 2025Build

    First commit, hackathon-style

    Kenil’s first commit was “Full Video Extraction.” 87 commits followed in October alone.

  3. 12 Oct 2025Build

    A frontend on the edge

    The web app went onto Cloudflare Workers from day one.

  4. Nov 2025Ship

    Payments and a cleanup cron

    Razorpay orders with a webhook backup, and a job that expires free-tier footage after seven days to keep costs bounded.

  5. Dec 2025Ship

    B-roll in the render; a real landing page

    B-roll composited server-side, project management, and the pricing, features and support pages.

  6. 9 Jan 2026Tell people

    Priced for India and the world

    Location-based pricing: INR at home, USD everywhere else.

  7. 26 Apr 2026Build

    Exports move to a worker

    Rendering left the request path for Celery and Redis — with timeouts, retries and progress steps the editor can poll.

  8. May 2026Research

    39 experiments

    A programme of standalone apps to prove an editor can run entirely in the browser: decode, composite, cache, mix, export, transcribe.

  9. Sep 2026Ship

    Stock, sound and email

    A stock library across four providers, audio layers, export emails and a free-tier watermark.

  10. BuildNext

    The agentic layer

    On-device silence cutting, auto-reframing and smart-cut ranking — the parts of an editor that edit for you — moving from experiments into the product.

Built with

  • Python · FastAPI
  • Celery + Redis
  • FFmpeg + libass
  • Next.js on Cloudflare
  • WebCodecs · WebGPU
  • Razorpay · PostHog

Writing about ReelForge