Systems

I also build the machines that make the work.

Most AI creatives deliver assets. Alongside the art direction, I build the pipelines that produce them every day, with automated checks that catch a wrong number, a clipped line or text over a face before anyone else sees it. Three of them below, each with figures from its own run logs.

59short-form reels delivered by the reel engine
46ad sizes from one Photoshop file, in about 10 minutes
336book pages laid out unattended in about 3.2 hours
13.7Mviews on my top five campaign reels →
SYSTEM 1

A daily reel engine that costs nothing per reel

Python, Remotion (React), ffmpeg, a local Hindi voice-clone model · 2026 · runs my own pages daily

The problem

A page that posts every day needs a finished, on-brand reel every day. Editing by hand eats hours, and AI video credits charge per clip, so volume gets expensive fast.

What I built

A script goes in; a finished 9:16 reel comes out. The voiceover is generated locally with a cloned Hindi voice, cards and captions render in Remotion, and ffmpeg assembles the cut. Depending on the format, the build refuses to finish if:

  • runtime outside the format's band, or any beat too short or too long
  • loudness away from −14 LUFS, measured on the delivered file
  • the presenter off screen for more than 12 seconds
  • a news clip without an on-screen source credit
  • hook text inside the measured head area of the clip under it
  • a pronunciation check (speech read back and compared) below its threshold

One engine now serves any brand: a client is a folder with a logo, colours, presenter clips and a voice sample. I proved that refactor changed nothing on the existing pages by re-rendering 205 frozen frames and diffing them pixel by pixel: zero difference on every frame.

59
reels delivered since August 2026
45–55 min
for two finished reels, research to file (measured 17–18 Sep)
0
generation credits per reel on the daily path
205 / 205
frames identical after the multi-brand refactor
Coaching brand reel frame: presenter under the title 3 galtiyan, kam marks
Coaching
News desk reel frame: sourced footage card about Census 2027
News desk
Campaign desk reel frame: a ledger of dumpsite figures with sources
Campaign desk

Three brands from the same engine. The brands are samples; the pipeline and figures are real.


SYSTEM 2

46 ad sizes from one Photoshop file

Python, psd-tools, generated Photoshop scripts · 2026 · built for my team's display campaigns

1080×2592 1080×1814 1920×1080 1080×1080 1600×1050 1440×864 1440×720 1349×657 600×600 990×505 978×487 1440×480 640×480 1168×450 1000×450 1160×440 1120×400 1080×400 970×400 1080×250 970×250 980×200 728×180

The problem

Every display campaign needs the same creative in dozens of Google Display sizes, from 320×50 strips to full-screen 1080×1920, and each one gets rebuilt by hand.

What I built

The engine reads the master file's layers, works out what each one is (background, subject, headline, button, logo), plans a layout for each size family, and writes the Photoshop scripts that build and export every size. Each size is left as an editable layer group, so a designer finishes the last 10% instead of starting from zero. Hindi headlines stay live, editable text.

QC checks every export's dimensions and file name, flags blank or near-blank exports by pixel variance, and produces a contact sheet for review.

46
sizes exported on one campaign run
≈10 min
from first to last export on that run
4
campaigns built with it
editable
every size stays a layer group with live text

SYSTEM 3

A 336-page report, laid out unattended

Python, CorelDRAW automation (COM), python-docx · 2026 · built for a print report

  1. Read the templateType sizes, colours and roles measured from the designer's file
  2. Parse the manuscriptWord file read verbatim, with an integrity check on every block
  3. Plan the flowsHeadings, section openers, figures and 228 tables placed in order
  4. Build the pagesText poured into frames with real overflow measured; long tables split across pages
  5. Run QCNine gates: rendered ink, word-level copy match, fonts and type sizes
  6. DeliverEditable master, press PDF and a review PDF

The problem

A long report in a house template, where every word has to match the approved manuscript and a single dropped line is a correction after print.

The part that mattered

The QC gates, not the layout. The system checks the copy word by word against the source and reports what it could not fix instead of hiding it: the one accepted blemish (29 paragraphs ending on a single word, about one per 12 pages) is written into the delivery note with the reason and the fix.

336
pages built
≈3.2 h
unattended build time
25 / 25
figures placed, none failed
9
QC gates run on the built file

comfyui-ad-resize

Open-source ComfyUI nodes that turn one layered ad into every display-ad size, with QC gates for blank output, readability and overlap. A clean rebuild of the ideas in the ad-resize engine above. github.com/aayush0124/comfyui-ad-resize

reel-qc

Open-source quality gate for short-form video: loudness, black and frozen frames, silence and a static opening, with exit codes a pipeline can act on. github.com/aayush0124/reel-qc

Book publishing pipeline

Manuscript to Kindle and paperback files, covers and metadata: 24 books live, royalties paid in five Amazon stores.

ClipTube

Desktop app, public on GitHub, that downloads exact time ranges of YouTube videos, for Windows and Mac. github.com/aayush0124/cliptube

Campaign art direction

13.7M views across my top five reels for a 9.4M-follower national account. Read the case study

Need the work and the machine that makes it?

aayush.paliwal01@gmail.com