Generative AI Pipeline Engineer

Local AI pipelines for documentaries, music channels, online stores and AI characters.

Script, image, voice, video, publish. I build the whole chain in Python and run it on one GPU, every day.

33videos published in four weeks
115klines of Python in my agent
9,316products in the store I built
16 GBof VRAM for all of it

Problems I take off your plate.

Content at volume, still on brand

A pipeline from script to published video: research, images, voice, music, edit, upload. Ten videos a week from one queue.

A character who stays the same

Fixed face, body, room and props across photos and video, with references, LoRA training and reference-driven video models.

Product visuals from a packshot

Scenes generated around the real product, with the label and shape kept intact, checked frame by frame.

ComfyUI workflows that run unattended

Graphs driven from Python, a GPU queue, retries, and gates that reject bad frames before a human sees them.

Local LLM agents and tool calling

Agents on Ollama models with tool routing, guardrails and benchmarks, so a cheap model still gets the job done.

Off the cloud bill

Everything above runs on a consumer card: no per-image fees, no client material leaving the machine.

Generation

ComfyUIZ-ImageKrea 2Qwen-Image EditMiniMax H3LTX VideoLoRA trainingYuE musicTTS and voice cloningWhisper

Code

PythonFastAPIFlaskaiogram 3SQLAlchemyPlaywrightffmpegpytestElectron

LLM

Ollamagemma4function callingRAG on Chromaevals and benchmarksClaude Code

Ops and APIs

LinuxOracle CloudsystemdCaddyGitHubYouTube Data APIMeta Graph APITelegram Bot API

Craft

After EffectsPremiere ProPhotoshopFigmaten years of motion and video

Three channels, one goal: monetisation.

An experiment I run in public. Niches picked from data, a pipeline built for each, decisions taken from YouTube Analytics. The Partner Program threshold is not reached yet.

Ages LivedHistory documentaries

Seven films of 15 to 23 minutes about ordinary life in the past. Script from checked sources, flat illustrated style, my own fact gate.

Side A NightsSad deep house mixes

Eight mixes of 36 to 60 minutes. Every song written, sung and mastered by models, every cover a seamless animated loop.

Cover: It wasn't white Cover: The night Rome burned Cover: First strike in history Cover: Not warriors

Plus 18 horror shorts on a third channel. Weekly output is about ten videos, produced by the same queue on the same card.

One episode, eight commands.

How What Did Medieval Peasants Actually Do All Day? was made. Every number below comes from that run's logs.

research

12 sources, one myth caught

My agent collects facts with citations. A gate flags weak claims before the script: they drank ale because water was unsafe was stopped here.

script

4,287 words, 15 chapters

Narration written chapter by chapter. Paragraphs become the beats the storyboard splits into scenes.

storyboard

131 scenes as data

{
  "id": 10,
  "chapter": "Before dawn",
  "cast": ["MOTHER"],
  "subject": "kneeling by the hearth,
    spreading ash over embers",
  "time_of_day": "night",
  "seconds": 12.0
}
Scene 10: a mother covers the embers for the night illustrate

131 frames, about 23 s each

Z-Image Turbo with a flat style LoRA, the same characters in every scene.

narrate

Word error 1.6 to 4.3%

Each chapter is voiced, transcribed back and compared with the script. A chapter above the limit is redone automatically.

score

Music with zero words

An instrumental is generated per episode. A check rejects any take where the model started singing.

edit

23.4 min at -14.1 LUFS

Clips, subtitles, chapters and loudness assembled by code. Items left for a human to review: zero.

Final cover: They slept with cows publish

Scheduled through the API

Title, chapters, sources, cover and AI disclosure set by the uploader, then it goes live on time.

Own product, since May 2026

Miks, the agent that runs it.

A desktop assistant on local models. It researches, writes code, runs the tools and checks its own answers. Every gate in it came from a real failure and is locked by a test.

115k
lines of Python
2,300+
tests
37/37
tool-router benchmark
19k to 1.5k
tokens per request
llm-tool-routerThe router, published on its own: benchmark, 24 tests, MIT.
  1. RequestRussian or English, voice or text
  2. Routerpicks 9 to 12 of 100+ tools by the user's own words
  3. ReAct loopruns tools, executes generated code, retries on failure
  4. Gatesreject answers that claim what the data never said
  5. Answerwith the files it actually opened
Miks desktop app: a question answered after two file reads, with the tool ledger on the right
A real run on a local gemma4 12B: two file reads, a correct count, 6.2 s. The ledger on the right logs every tool call the agent has made.

Vibe coding, with guardrails.

Most of my code is written together with AI coding agents: Claude Code and my own Miks. I set the goal, review every diff and measure the result. Tests decide what ships.

648 of 737

commits in Miks made in a pair with an AI agent. The other 89 by hand.

2,300+

tests and a pre-commit hook. An agent's change that breaks one does not land.

19 vs 23

tool calls out of 35: a Russian prompt cost a small model real calls. Measured on a bench, then fixed.

20

design decisions written down as notes, so a rejected idea never gets proposed twice.

This page was built the same way: one afternoon, an agent at the keyboard, me deciding and checking every screen.

Kiss and Tell Mini App home screen in Telegram
Generated scene: the book on a bedside table by candlelight The supplier's original packshot of the same book
Drag: supplier packshot, then the scene generated around it.
Generated mood shot: coffee and a book on a bed

A store that lives in social feeds.

Kiss & Tell, an adult wellness brand in Ukraine (18+). I built it alone: the shop, the content and the posting.

  • Telegram Mini App9,316 products from a supplier export, cart, checkout, admin with click tracking. Runs 24/7 on Oracle Cloud.
  • Generated contentProduct photos and reels made locally from supplier shots, every frame checked by eye.
  • AutopostingThree posts a day to Instagram, Threads and Facebook through the official Meta APIs.

Threads views per post

One-line question117 to 460
Product post5 to 16

So the feed moved to conversation first. Click tracking then showed the next gap: the audience grows, store visits do not yet.

Kira, a face that stays the same.

The store's brand ambassador is an AI character I designed: her face, body, apartment and even her phone are fixed references. She posts on Threads in her own voice and stars in the reels.

  • Reference-driven videoMiniMax H3 with up to nine reference images: face, room, single garments. The wall and baseboard match to 2-3 px across separately rendered clips.
  • Upscaled in latent spaceDrafted at 576x1024, upscaled to 1088x1920 in a second pass. Fits on the same 16 GB card.
  • Her own LoRA, measured firstTraining on Qwen-Image peaks at 14.7 GB, so it runs locally without offloading.
Kira at the window of her apartment at night Kira in a cafe with two cups of cocoa Kira's evening skincare selfie Kira working on a laptop Kira with a cup of tea on the sofa

Before the pipelines.

2025 - now
Generative AI and automation engineer

Own projects: video pipelines, Miks, the store, the channels.

2023 - 2026
Creative designer, affiliate marketing company

Performance creatives for Facebook Ads in a media-buying team, in-house and remote. Early AI video work.

2022 - now
Motion designer, freelance

Logo animation, After Effects templates, ad video. Earlier: motion, video and sound since 2017.