Punchbox — a dictation, transcription and vertical-video studio, designed and run by Wolf Agency.
This is an engineering demonstration, not a sales pitch. Punchbox is a real product Wolf builds and operates: you speak, it writes, you proofread, it can read the text back and build a subtitled vertical video. It is deployed and running on punchbox.fr.
Honesty first: Punchbox is in pre-production. The free trial works, but paid subscriptions are not open yet — so there are no clients, no revenue and no testimonials to report. Everything on this page is backed by a real trace: a mission report, a commit, or a measurement.
Punchbox runs on its own domain, served from an encrypted volume behind an nginx edge with TLS. The free trial — 30 minutes of audio, no card — works today; paid checkout is not open yet.
punchbox.frPunchbox is one tool with one job: turn what you say into text, and text into a vertical video ready to post. It runs in the browser as an installable app, with no third-party API involved.
You speak; Punchbox transcribes; you proofread; it can read the text back and build a subtitled vertical video. The client posts it themselves — nothing is hosted or recommended, which is exactly what keeps a one-person product legally and ethically clean.
Every engine runs locally and frozen on Wolf’s own server: faster-whisper for transcription, Piper for speech, ffmpeg for video. There is no third party to call — so there is no one to leak to.
Speak into the browser; faster-whisper (large-v3-turbo) transcribes across 27 languages, with automatic language detection.
The transcript lands in an editing rack — correct it, translate it, prepare it before it becomes a video.
Text becomes a synthetic voice: 13 voices across 8 European languages, every voice licence traced to its source.
Nine installed translation pairs, each with a verified licence, so a transcript can be rendered in another language.
A subtitled 9:16, 16:9 or 1:1 video is built from your own audio or a synthetic voice — no GPU needed.
Installable as an app on Windows, Linux and Android; source audio is auto-deleted within the hour; nothing is sold or shared.
Reserved slots for screenshots of the running product. Real captures drop in without touching the layout.
The recording console — transport, level meter and live text.
The record button and the VU-style level metering.
The transcript, with source and translation side by side.
The vertical-video builder and its settings.
The compact floating panel with full transport.
The Punchbox charter and the legal pages.
A non-exhaustive list of what Wolf actually produced on Punchbox — each item backed by a report, a commit or a measurement.
A closed design system in three formal charters — graphic, interface and density — a brass-on-bakelite radio-console register, a single stylesheet, and zero third-party resource.
Every repository and deliverable is anchored with OpenTimestamps (SHA-256 → Bitcoin): the hash is public, the document stays private, and anyone can verify it.
A 20-language interface with full dictionary parity — the same keys everywhere, none missing — including Arabic and Hebrew right-to-left via CSS logical properties.
faster-whisper large-v3-turbo covering 27 languages with automatic detection; measured at about 4× real time on CPU, 1.66 GB of RAM per job.
13 Piper voices across 8 European languages; each voice traced to its corpus and licence (CC BY / CC BY-SA), with the mandatory attribution carried in-product.
Nine translation pairs installed, each licence verified, so a transcript can cross from one language to another.
A prompter that renders 9:16, 16:9 and 1:1 video; a one-minute vertical clip costs about 7 seconds of compute — measured — with no GPU.
A .srt subtitle track is generated and aligned to the voice.
Client data sits on a LUKS2 volume (AES-XTS, 512-bit) that mounts automatically at boot — proven with the account database byte-identical before and after.
Source audio is auto-deleted within the hour by a systemd timer; IP addresses are never stored in clear (salted HMAC-SHA256, bounded retention); the client picks a retention window capped at 30 days.
Off-site, encrypted, deduplicated backups (restic to a 1 TB Storage Box); a real restore was executed from off-site and verified — dump integrity checked, 37 tables read back.
WCAG 2.1 contrast computed pair by pair by script: body text at 16.26:1 (AAA); accent-as-text uses flagged at 2.07:1 below AA, then corrected to zero pairs below threshold in both themes.
How the pieces fit: an nginx edge in front of an installable app, local speech engines, an encrypted data volume, and a delivery chain that proves itself.
An nginx edge terminates TLS (Let’s Encrypt) and serves punchbox.fr.
An installable progressive web app — HTML/CSS/JS, no framework, no build — with versioned, cache-busted assets.
A FastAPI service with a strict single-worker queue, key auth, quotas and GDPR purge — parallel workers were measured to add nothing.
faster-whisper and Piper run locally and frozen on Wolf’s server; no third-party API is ever called.
Client data lives on an auto-mounting LUKS2 volume, with a systemd timer purging source audio within the hour.
A deploy script that proves the container actually restarted — freshness within 120 s, image-ID match — and refuses to claim success otherwise. It once honestly reported a deploy as NOT PROVEN and stopped.
Nightly off-site encrypted backups (restic, restore tested), plus OpenTimestamps anchoring of every repository and deliverable.
Two things Wolf measured rather than claimed: how fast the engines run, and how the interface scores against WCAG. The numbers come from executed benchmarks and a pair-by-pair contrast script.
Benchmarks run on the production CPU (4 cores), three concordant passes — no GPU anywhere in the product.
Because the models run locally and never change, training on your voice is impossible, not merely forbidden. There is no third party to call, so there is no one to leak to.
Contrast ratios computed by script on the colour pairs actually used — the good and the bad reported alike.
Every claim on this page is tied to a trace. The engineering ran as tracked missions, each ending in a report with executed proof rather than a code reading; benchmarks were run three times; the interface audit ships its script and its formula.
Punchbox is how Wolf works, made visible: design, measured accessibility, industrialised deployment with restart proof, proven backups, GDPR and timestamped anteriority — on a product that actually runs.
The same method applies to other projects under the same terms. Punchbox is in pre-production; there are no client or revenue figures to show, and we will not invent any.