Work · Zeely AI · December 2025 — June 2026
For half a year, Flavatars’ founder worked inside Zeely AI — the platform that gives small businesses ads at big-brand level — turning generative video from one-off luck into an ad production engine.
The production line · tap any ad for sound
In short
Inside Zeely AI, Flavatars turned generative video into an ad production system. Half a year embedded in the product team as AI Video Production & Workflow Architect: realism rules that strip the AI gloss, reference-true cloning, hook craft, voice-over films, multishot continuity, batch variants. The proof plays on this page — the ads themselves, with sound.
The origin
Before any contract, there was a trial: make something that proves the level. This vertical promo was that answer — cinema pacing in an ad’s body, built alone, delivered fast.
Zeely signed. What followed became this case: half a year of building the systems behind ads like the ones above.
The client
Zeely AI is one of the boldest products in performance marketing: an AI platform where a small seller gets what only big brands used to afford — strategy, ad creative and campaign launch, run end to end. More than two million people use it, and for many first-time entrepreneurs it is the first AI product they ever touch.
It is also a rare kind of team: product engineers, marketers and creators shipping at a speed most studios cannot even brief at. Work made there in the morning meets real users by night. For a creative craftsman, that pace is the honest test — and the reason this engagement became a system, not a folder of videos.
The ask was direct: make generative ads believable, and make believable repeatable.
The system
The method reads like a director’s discipline written into software. Six moves, each one a craft of its own:
The reference ad is taken apart shot by shot: what happens, why it works, what job every frame does in the story.
Visual Function Fidelity — a shot’s purpose (effort, proof, payoff) survives the move to a new product even when nothing on screen stays the same.
Character, wardrobe, room, light — described so precisely that any single frame can be rebuilt on its own and still belong.
Video instructions written like a director’s sheet: staging, camera grammar, speech timing, where the eyes look, what stays quiet.
Films are generated in connected blocks, not stitched-together shots — so the voice, the face and the room survive every cut.
Every output is rated against the reference. The strongest takes become the new bar; the weak ones teach the rules.
The point of a system is that a strong result is never a lucky one-off. Two runs on the same Crystal Mind brief — different performances, one discipline. Tap either for sound.
From the floor
One heroine kept turning into a different woman between shots. Swapping models did nothing — version after version, the drift stayed. The answer was never the model: it was how her reference travelled through the system. One architectural change, and she finally stayed herself. Engineering thinking beats tool-hopping.
One morning the analysis service simply refused to work, and the day’s production stood still. By evening Oleg had assembled his own video deconstructor — footage cut into shots, speech transcribed, work moving again. The craftsman had quietly become an engineer.
The final test was against his own hands: creatives made the manual way versus the system’s output — first takes, no cherry-picking, no fixes. He wrote the fairness rules himself. His own craft became the bar for the machine he built.
Disciplines
The ad cannot look like an ad. iPhone framing, real rooms, imperfect light — a written set of rules strips away the AI gloss shot by shot.
A winning ad is a structure, not a file. Its pacing, framing and story carry over to a completely different product — legally and recognisably new.
The first seconds decide whether the rest exists. Hooks are studied, catalogued and rebuilt product-native — energy preserved, story intact.
A written script becomes a performed film. The voice keeps its exact words; the picture learns when to cut, breathe and land with it.
One person, one wardrobe, one home — across every cut. Films are generated in connected blocks so nothing drifts between shots.
One product in — a family of creatives out: different presenters, different reads, the same discipline. This AG1 take is the sister of the one above.
The same face stays the same face. Character identity is engineered into the system, not hoped for — frame one to frame last.
Room tone, footsteps, fabric, a box opening. Believability is audible — the films carry the small sounds a phone would honestly record.
Series artwork
Poster artwork for Flavatars × Zeely — the engine’s disciplines, one word at a time.
The arc
01
Watch everything first. Name what holds the creative back, fix the highest-value thing before touching anything else.
02
Lucky prompts become repeatable looks: a described visual language a whole team can shoot in — describing the result, not the process.
03
Presenters built once and reused everywhere: faces, wardrobe and manner that survive from ad to ad.
04
Why the first seconds win — studied across live performance ads, then written down as rules a system can follow.
05
The pieces connect: from reference breakdown to a finished film, one system with quality gates instead of a chain of hand-offs.
06
Real briefs, real product categories, and an honest test of system against hand craft — first takes on the table, no favourites.
What the client keeps
Facts
| Client | Zeely AI — AI ad creative platform (Zeely, Inc., USA) |
|---|---|
| Period | December 2025 — June 2026, embedded in the product team |
| Role | AI Video Production & Workflow Architect — Oleg Pylypenko |
| Mandate | Turn generative video into a repeatable ad production system |
| Disciplines | UGC realism · reference cloning · hooks · voice-over · multishot · batch variants · identity control · sound |
| Formats | Vertical UGC video ads, hook-led creatives, voice-over films |
| Models in the loop | Gemini · Claude · Kling · Seedance · Veo · Higgsfield |
| Showcase | The ads on this page — original soundtracks, one click away |
| Handover | Working systems, creative rulebooks and documentation the team builds on |
FAQ
For half a year Flavatars’ founder Oleg Pylypenko worked inside Zeely AI as AI Video Production & Workflow Architect: he designed the creative rules and production workflows that turn generative video into repeatable, believable UGC advertising — from reference breakdown and hooks to voice-over films and batch output.
A performance ad that looks like real creator content — phone framing, natural speech, everyday rooms — but is produced with generative AI. Done right, the viewer reads it as a person’s honest video, not a commercial; that credibility is what makes the format perform.
Generative models default to cinema: glossy light, perfect skin, dramatic lenses. The system fights that with written realism rules — phone-native framing, imperfect light, real environments, believable pacing, ambient sound — applied to every shot, so the gloss never comes back.
Yes — if production is a system rather than a series of lucky generations. Structured creative rules, connected generation steps, identity control and honest evaluation turn one brief into families of consistent ad variants across products and presenters.
Carrying over what makes a winning ad work — its dramatic structure, pacing, shot logic and the job each frame does — to a different product, without copying the footage. The result is a legally new creative that keeps the reference’s persuasive engineering.
Zeely AI is a USA-based AI ad creative platform used by more than two million people: small businesses get strategy, ad creative and campaign launches at a level that used to belong only to big brands. For many first-time entrepreneurs it is the first AI product they ever touch.
The studio
Flavatars builds shows — and builds the machinery behind them. In our own portfolio, Zeely stands beside the digital twin we built for BoConcept’s owner and the interactive CEO avatar we built for Huawei — different products, one discipline: creative direction translated into systems that keep working after we leave the room.
Great ads aren’t generated. They’re engineered.
We design AI video production systems end to end — creative rules, workflows, quality gates — and hand them over working.
