AI automation
Three automations and their numbers
I don't write code by hand. I define the job, direct AI coding agents (mainly Claude Code) to build the tool, test it on real files and measure the result. I keep what runs and I hand over the files and the instructions.
The three cases are my own projects, not client work. Every number comes from the files on my disk.
Updated October 26, 2026
Case 1
20 educational Shorts made by one pipeline
The problem
For my own educational YouTube channel, every Short needed manual animation, a voice track, a music mix, a loudness check, a loop check and an upload sheet with title, description and date.
What I built
A pipeline that takes a character, a short story and a sentence. It returns:
- a 50.5-second vertical video;
- a contact sheet and a cover;
- loudness and loop measurements;
- the YouTube title, description and publish date;
- a copy under 10 MB for browser upload.
I picked the AI voice in a blind test of three engines (CosyVoice3, IndexTTS-2.5 and Qwen3-TTS) and checked it with Whisper: 143 of 147 syllables right.
The result
- 20 of 20 videos passed the automatic checks: 50.50 s each, between −15.8 and −16.4 LUFS.
- 16 were produced in a single day, at about 3.3 minutes of render per video.
- All 20 are scheduled on YouTube, Tuesday, Thursday and Saturday at 6 p.m., from October 20 to December 3, 2026.
Tools: Node.js, headless Chrome, ffmpeg, Python, Whisper, CosyVoice3, YouTube Studio
Case 2
Choosing which pages to write from 924 real searches
The problem
My own jewelry business (ivebeen3d) was guessing which pages to write for its website.
What I built
A pipeline that starts from 34 seed terms and:
- collects what people in Colombia actually search;
- records the top 5 results for each search;
- groups searches that share the same results into topics;
- collects competitor prices;
- drafts one page per topic.
The result
- 924 searches collected, 911 measured, grouped into 531 topics.
- 491 competitor prices.
- 12 topics recommended.
- 6 pages put live on ivebeen3d.com on October 19, 2026: 4 new and 2 rewritten. All 6 were submitted to Bing through IndexNow.
The six pages, in Spanish: ring size (rewritten), silver care (rewritten), graduation rings (new), custom pendants (new), ordering a custom ring (new), wedding bands (new).
Tools: Python, Next.js on Vercel, IndexNow, Claude for the drafts
Case 3
Product photos from 3D files, in batches
The problem
My archive of thousands of 3D files had no images. The renders that did exist looked fake, and the stone settings were empty.
What I built
Two tools:
- a batch renderer that makes one 2K catalog image per file;
- a photo pipeline that sets the stones automatically and returns a 1600×2000 photo on velvet, a 1080×1350 crop for social media and a copy of the 3D file with the stones in place. The original file is never touched.
The result
- 3,004 catalog images in three days (August 5–7, 2026), from 2,180 STL and 615 3DM files.
- 104 realistic photos, chosen from a sweep of 22,054 3D files.
- 35 photos redone in one day after a correction to the prongs.
Tools: Python, Blender (Cycles), rhino3dm, numpy, Pillow
More
- Batch quotes for 3D printing: slices without opening the slicer and returns minutes and grams per plate. More than 50 jobs; one project of 7 plates came to 52.1 h and 1,268 g.
- Castability check for jewelry files: thickness map, weight in 4 metals and a verdict, in 1 to 4 minutes per piece.
- Batch conversion of Rhino 9 files to Rhino 8: 48 files and 6,607 objects checked, each with a backup.
- Daily website monitoring: 62 pages and 92 links checked; it only alerts when something breaks.
- Invoices with signed PDFs: for a corporate client.
Contact
Iván Suárez, in Bogotá, Colombia. For a project or a remote role, write to hola@ivebeen3d.com or find me on LinkedIn.