English and Malayalam
Shipped across 20 finished pieces, each with a human proofing pass.
V1Edit engine · English · Malayalam · moremultilingual · 100+ languages
Every edit is a set of decisions. We compute them.
EDITSON takes raw footage and returns a finished, on-brand film: the cuts, captions, B-roll, graphics and mix. It runs on our own NVIDIA RTX hardware, so long 4K footage never leaves the building.
20finished pieces, TikToks to podcasts
120minutes of delivered runtime
3language modes: English, Malayalam, bilingual
100+languages in our speech stack, 2 proven in production
2formats: 9:16 shorts and 16:9 long-form
A1The stack
Built on NVIDIA. Rendered on RTX.
CUDA runs the speech models and matting. NVENC encodes the final cut. We work with NVIDIA and AMD, and our content is built around the hardware both make.
PartnerNVIDIA
CUDA pipelines and NVENC rendering on GeForce RTX.
PartnerAMD
Ryzen and Radeon builds are a core part of our content.
GPU workstation24 GB VRAM
A flagship-class CUDA GPU with 24 GB of video memory for long 4K sessions, on site.
PipelineCUDA + NVENC
GPU speech recognition, matting and hardware video encoding.
Partnership wording and logos are subject to each partner's brand guidelines.
T1Algorithmic editing
Drag the threshold. Watch the edit change.
An edit starts with a question: which pauses are dead air and which are timing? We measure the audio and answer it with numbers. Move the slider to see how one rule reshapes a take.
speech keptshort pause keptpause cut
Illustrative signal, generated in your browser. The pipeline applies the same rule to real voice tracks, with extra checks so words are never clipped.
60.0soriginal take
60.0safter cuts
0pauses cut
VOICE
Energy, pitch range and pace pick the best take of every line.
FACE
Face and lip tracking keep the speaker framed and sync extra cameras.
WORDS
Word-level timing puts every caption on the exact moment it is spoken.
V1The pipeline
Six stages from raw footage to post-ready.
Each stage is its own module, so a partner can adopt one stage or the whole line. Select a stage to see what it does.
T1Visual styles
One engine. Any look.
Style is data: fonts, palette, caption behaviour and how graphics enter. We load a new look per brand without touching the cutting logic.
A2Languages
One engine. Over 100 languages.
Our speech stack transcribes more than 100 languages, including 25 Indian languages with a dedicated model. Two are proven in production today. Every other language is benchmarked on your footage in a pilot before we commit to it.
Indian languages
A dedicated Indic speech model covers the 22 constitutionally recognised languages, plus English, Bhojpuri and Bhili.
Global languages
A general multilingual model covers Spanish, French, German, Portuguese, Arabic, Mandarin, Japanese, Korean and 90 more.
provendedicated modelbroad model
No match. Ask us: new languages are added per project.
Support means automatic transcription plus subtitles with a human proofing pass. Accuracy differs by language, and we measure it per language before delivery.
V2The work
Fourteen clips. Press play on any.
Excerpts from finished videos for real shoots. Everyone shown has given consent for pitch use. Drag or flick the row, use the arrows, or tap a clip for sound.
A2Build log
Shipping pace, in October 2026.
Dates come from our project files and are rounded to the day. Each entry is a finished deliverable, not a prototype.
The house style is written down
Captions, cards, grade, music and sound rules become a style guide the pipeline follows on every video.
English TikToks: $1,000 PC and a second build
Word-by-word captions, spec cards, cutouts over B-roll and a ducked music bed.
Malayalam and bilingual reelsMultilingual and bilingual reels
A Malayalam speech stack, English subtitles and a second visual style, the neon build reel.A multilingual speech stack, English subtitles and a second visual style, the neon build reel.
Two YouTube films, about 15 minutes each
Chapter labels, spec cards, live FPS overlays, thumbnail and post kit.
Two podcast episodes and nine branded reels
A 71-minute 4K two-camera session cut into two episodes, plus nine client-branded reels from a separate studio shoot.
Tap for sound
V1Why it is hard to copy
Owned compute. Indic fluency. Taste, encoded.Owned compute. Many languages. Taste, encoded.
Local GPU
Long 4K, multi-camera sessions run on our own RTX workstation with no upload caps, and client footage stays private.
Malayalam, handled
On our benchmark, the best model read about 80% of Malayalam chunks correctly where stock Whisper failed. A proof sheet covers the rest.
Many languages, one pipeline
Our speech stack covers 100+ languages, including 25 Indian languages with a dedicated model. Two are proven in production; every other language is benchmarked on your footage first.
Taste, in code
Pacing, framing, card timing and loudness come from rules refined across twenty finished pieces.
Benchmark: one two-minute Malayalam and English test recording, split into chunks of seven seconds or less.Accuracy differs by language. We measure it on your footage in a pilot before committing to a language.
T1Roadmap
From editing to a multimedia content engine.
NOW
Video finishing
- Shorts, YouTube films, podcasts
- English and MalayalamEnglish and Indian-language content
- Branded series
NEXT
Productise
- One-command runs
- Client review interface
- Brand style library
- More Indic languagesMore languages, verified one by one
LATER
Multimedia
- One recording, ten assets
- Audio-first podcast edits
- Localised versions
- Platform integrations
Only the first column ships today. The rest is plan.
A1For investors and partners
Four ways to work with us.
A
White-label production
We finish your clients' content under your brand.
B
Platform integration
Our engine behind your podcast or editing product.
C
Pilot project
Send one real episode. We return a finished cut to judge.
D
Investment
Fund the step from proven workflow to product. Terms to be agreed.
Where we stand: the output is proven. The per-project scripts still need to become one reusable product, and parts of face tracking and rendering still run on CPU. That is what a partner unlocks.
V1Next cut
Send us one episode.
We return a finished, post-ready cut so you can judge the work.