S.Ler Group — AI Technology Company

Research — NotebookLM

An 11-section structured interview protocol runs over 50+ peer-sourced citations per topic (PMC, USDA, NIMH). Three phases: baseline → deep dive → synthesis, with a blind-spot pass and manual retry for hallucination-prone questions. Output: ~43k characters of source-backed research per episode.

Script — Claude Opus + custom playbook

Channel-specific doctrine files enforce voice, cadence, and narrative arc. Reverse-countdown structure (10 items → killer #1), seven-step opening hook formula. Output: ~120 segments × 15-17 words = 9-10 min target runtime at 207-210 wpm. Eight-pass internal quality loop before sign-off.

Visual prompts — GPT-5.4 chunked

122 cinematic-storyboard prompts per episode, generated in 2×61 chunks (single-pass times out). Per-channel visual grammar bakes in palette, hero composition, anti-AI keywords, material-specific detail. Fully automated, no manual prompt editing per episode.

Images — FLUX-fp8 + custom-trained LoRA

Four character LoRAs trained in-house — Frank, Maggie, Dave, Blob — one per persona. FLUX.1-dev-fp8 base, CFG 3.5, 20-30 steps. Renders on Spot A100 80GB at ~2 sec/image, 122 images per episode. Character identity holds across all 122 frames via LoRA conditioning, not per-shot face-swap.

Motion — LTX-Video 13B Distilled FP8

First segment of every episode is a 3-keyframe image-to-video stitch — a ~15-second motion hook to retain audience past the 8-second drop-off. Speed-matched to the TTS waveform so cuts land on the beat.

Voice — Qwen3-TTS-12Hz-1.7B + custom clones

A 30-60 second reference per channel produces a stable voice clone. 122 segments are TTS'd individually, gap-trimmed, sequenced — no monolithic generation, because per-segment retry is cheaper than re-rolling the whole episode.

Assembly + publish — FFmpeg + tooling

FFmpeg-driven assembly with SFX (riser cues, beat-aligned punches at 120 BPM, -18 LUFS body / -14 LUFS punches). A vertical 9:16 short — 14 scenes, seamless loop hook, 30s — is generated from the same source. Three thumbnail variants. Manual upload + community post 24h before publish.

Built on Google Cloud

Every image, every video frame, every voice clip is rendered on Google Cloud GPUs. Compute Engine with L4 for fast iteration, Spot A100 for batch production. NotebookLM for source-backed research. The entire pipeline lives in one Google Cloud project.

Compute Engine — 2× A100 80GB Spot VMs at production cadence, ~400 GPU-hours/month

Cloud Storage — ~500 GB media archive, version-controlled, lifecycle-tiered

NotebookLM — source-backed research, 50+ citations per topic

Vertex AI — pipeline orchestration, batch jobs, model serving for in-house LoRAs

IAM + Quotas — fine-grained cost and resource control across the project

Target monthly burn at production cadence

~$500–600 / mo

GPU compute (~$480) + storage (~$10) + Vertex AI orchestration. Credits would cover ~18–24 months of operations at the target cadence — directly funding the scaling from 5 to 10+ channels.

We build AI infrastructure for video content production — and deploy it across 5 live YouTube channels.

Five live deployments. One pipeline.

Frugal Forward

The Lazy Grocer

Inner Mechanics

Lived

Rhymes of History

Five recent episodes. One per channel. All produced by the pipeline.

Her $42K Paycheck Cost Us $4,000. Two-Income Trap.

Dollar Tree's NEW $1.50 Garden Products — What Actually Works (and 3 to Skip)

Ferrari Brain. ADHD.

Rasputin — The Night He Wouldn't Die

You Are on Dyatlov Pass: From Nightfall to Dawn

How one 10-minute episode actually gets made.

Research — NotebookLM

Script — Claude Opus + custom playbook

Visual prompts — GPT-5.4 chunked

Images — FLUX-fp8 + custom-trained LoRA

Motion — LTX-Video 13B Distilled FP8

Voice — Qwen3-TTS-12Hz-1.7B + custom clones

Assembly + publish — FFmpeg + tooling

Why this is genuinely difficult

Built on Google Cloud

From 5 channels to 10+ by end of 2026.

One operator. One vision. Built in the open.

Valerii Serko

Partnerships, credits, investment.