Each entry says what the job was, what was actually done, and where the numbers come from. Where something has a limitation, the limitation is written down too.
All of it is produced in-house as demo and spec work — these are not client commissions. Where a figure is quoted it is measured from the delivered file. No testimonial or client logo appears anywhere on this site, because I don't have any to show yet.
A long-form recording repurposed into vertical clips for Reels, TikTok and Shorts. Each clip starts and ends on a complete sentence, so nothing opens mid-word or stops halfway through a thought — the failure that gets automated clips scrolled past. Reframed with the speaker held in frame, captioned, and normalised so the whole set plays at a consistent loudness when posted in a row.
Word-level caption timing styled to stay legible over a moving shot. The transcript gets a second review pass before delivery, because machine transcription reliably mangles proper nouns and technical terms.
The same footage exported as 9:16, 1:1 and 4:5 in one pass, each safe-area checked per platform so nothing important lands under a UI element, with hook text burned in at the top and a cover frame pulled from the source.
Colour, contrast and brightness matched shot to shot so footage captured at different times and in different light reads as a single piece — the unglamorous work behind every finished edit. Then cut on the action, with a ducked music bed and a −16 LUFS master.
Script built around one message rather than a feature list, voiceover, word-level burned-in captions so the ad works with the sound off, a CC0 music bed mixed to −16 LUFS, and delivery in all three ratios. Full commercial rights on the output.
Hiss, level drift and loudness repaired and mastered — measured on the actual before/after files, not estimated. Hiss energy down 11 dB, level variance between sentences cut from 8.1 to 5.9 dB, loudness range tightened from 9.6 to 8.5 LU, and the master lifted from −27.0 to −16.5 LUFS with a −1.5 dBTP ceiling so platforms stop turning it down.
Three separate jobs: a person taken out of moving video with the background rebuilt behind them, a person removed from a still, and a new object placed, lit and shadowed into an existing photograph. Grain and motion matched so the repair doesn't announce itself, and the audio left completely untouched on the video work.
Soft footage restored to clean detail and compared as a magnified crop rather than a whole-frame claim, so the difference is judged where it actually shows. Frame interpolation is available for smoother motion — 24 to 48fps without the ghosting frame-doubling leaves behind.
Voiceover generated per line rather than one continuous take, so the pacing fits the edit, then mixed and level-verified. English, German, Spanish, French, Italian, Portuguese, Russian, Japanese, Korean and Chinese are round-trip verified; Serbian, Croatian and Bosnian get a by-ear review before delivery.
Rather than applying a house style, the target company's live site was inspected to extract its actual typography stack, exact palette values and tone of voice — then a seven-slide Instagram carousel, a YouTube thumbnail, a Story cover and a feed stat card were built inside that system. Self-initiated concept work, delivered as a board plus the individual assets.
A supplied photo cut out, relit and composited onto a studio background with a real contact shadow, then delivered as a consistent set. Label text is repaired or added where the original packaging is unreadable at size.
Multi-scene sequences built scene by scene: stills generated and art-directed to a consistent look, then assembled with motion, titles, narration and a music bed into a finished piece rather than a slideshow.
Ten miles of track laid by hand in a single day, 1869 — a record nobody has beaten since — and the workers who were written out of the story. Built from a sourced dossier rather than from memory: the Chinese workforce is spoken as a range with the payroll analysis noted separately, and the claims the sources don't support are left out instead of smoothed over. Archival photographs from the Library of Congress, the National Archives and university collections, public domain, credited on screen.
Four hundred years of plans, a French attempt that ended in a graveyard and a scandal, and the American one that succeeded because it put sanitation before excavation and chose locks over a sea-level cut. The disease numbers and the money are spoken as the record gives them, not as a dramatised figure.
A tower put up for the 1889 exposition and meant to be dismantled within twenty years, kept alive because it turned out to be useful. Assembled from 1887–1889 construction photographs by Durandelle and other public-domain archives, with the work shown as it actually progressed rather than staged as a period look.
The three above are examples: new films go up on the channel regularly, and it also carries two pipeline demos (clip extraction, audio cleanup). Narration in each is synthetic and disclosed in the video description. There is no audience behind the channel yet — the films are published so the work can be inspected, not as a reach claim.
A note on honesty: the capability list on this site is deliberately shorter than it could be. Background removal as a standalone service, dubbing, custom music, standalone colour grading and green-screen keying are all left off — the pipeline for each isn't proven here yet. What's listed, works.