Skip to content
AI Engineering Field Kit

Video source

Eighteen short explainer videos covering everything an AI Field Engineer / Solutions Architect is expected to hold in their head, from prefill and decode through to multi-node serving. Built with Remotion 4.0.527 (React → MP4) and Kokoro-82M (Apache-2.0 TTS). No paid APIs anywhere.

Each video: a title card with what you’ll learn → concept scenes → worked examples with real numbers (before/after comparisons, sizing maths) → a “run it” scene with the actual commands → a three-point recap → an end card naming the lab to run next. Subtitles are burned in, in a reserved band that the diagrams never draw into.

# id covers pairs with lab
1 inference-101 prefill, decode, TTFT, ITL, goodput F1
2 kv-cache KV maths, PagedAttention, FP8 KV L5
3 gpu-bandwidth roofline, decode ceiling, device table L3 · L4
4 quantization FP16 → FP8 → 4-bit, KV quantization L4
5 batching continuous batching, chunked prefill, speculation L3
6 training SFT, LoRA, DPO, RFT — and how to choose F5 · L8
7 offloading sizing formula, offload order, MoE, TP vs PP L7
8 platform serving layers, capacity modes, maturity curve capstone
9 serving-stack what an inference server does; containers; the five flags L2 · L9
10 benchmarking traffic profiles, concurrency sweeps, percentiles L3
11 prefix-caching shared prefixes, RadixAttention, prompt caching L6 · F2
12 qlora frozen 4-bit base + adapter, data and evals L8
13 scale-out speculation, acceptance rate, two-box reality L9
14 full-finetuning training loop, 16 B/param, ZeRO/FSDP, when full FT pays T1 · T2
15 sft data format, loss masking, data volume, knobs, loss curves T2
16 peft LoRA B·A maths, QLoRA, DoRA, adapters, prefix tuning, IA³, multi-LoRA T2
17 rlhf preferences, reward model, PPO, KL leash, reward hacking, GRPO/RFT T1
18 dpo the DPO loss, one worked step, pairs from product data, IPO/KTO/ORPO T3

Total ≈ 57 minutes (videos 1–13 ≈ 2.5–3 min each; the training deep dives 14–18 ≈ 4 min each). Every end card names the repo lesson and lab to run next (field-engineer-labs/lessons/NN-*).

Regenerate on your Mac (MLX)

Section titled “Regenerate on your Mac (MLX)”
Terminal window
npm install
npx remotion browser ensure


pip install mlx-audio misaki soundfile numpy
python scripts/tts_mlx.py                      # all 18, voice af_heart, speed 0.96
python scripts/tts_mlx.py --video kv-cache --voice am_michael


npm run dev                                    # Remotion Studio
npm run render                                 # out/<id>.mp4

scripts/tts_patch.py --video <id> --scenes last re-voices only the scenes you edited. scripts/tts_onnx.py is the same script for a Linux/CPU box (kokoro-onnx + weights from the kokoro-onnx GitHub releases). It produced the delivered MP4s.

How it fits together

Section titled “How it fits together”
  • src/narration.json — single source of truth: every video, scene, visual and narration line. Both TTS scripts and the compositions read it.
  • scripts/tts_*.py — write public/audio/<id>/NN.wav (24 kHz) plus public/scenes/<id>.json with the measured duration of each scene, a 0.35 s lead-in and a 0.7–1.1 s tail for pacing.
  • src/Root.tsx — calculateMetadata reads those durations, so composition length always matches the real audio. With no audio present it falls back to a word-count estimate, so the Studio works before you run TTS.
  • src/Explainer.tsx — layout: header, visual stage, 200px subtitle band, progress bar. Subtitles come from src/lib/subtitles.ts, which splits narration into sentence-sized cues and times them by character count across the scene.
  • src/components/Visuals.tsx — every diagram. Each one takes p (scene progress 0→1) so the animation fills whatever length the narration turns out to be.
  • src/theme.ts — colours, fonts, fps (24).

Rendering notes

Section titled “Rendering notes”
  • Delivered files are 1248×702 (--scale=0.65) h264. Drop the scale flag for full 1920×1080.
  • npx remotion browser ensure downloads Chrome Headless Shell; on arm64 Linux only the headless shell is available, which is what Remotion uses anyway.
  • Remotion bundles FFmpeg — don’t install one.
  • Licence: Remotion is free for individuals and companies up to 3 people; CI rendering counts as an automation under their terms. Kokoro-82M is Apache-2.0.