Overview — Project 4: Cutroom

Cutroom analyzes a video without pretending one model can reliably “watch the whole thing.” Rust coordinates specialized stages: decode, sample, detect cuts, transcribe speech, read on-screen text, detect audio events, align a timeline, reason, and grade claims.

Product output

cutroom/run-01/
├── media-probe.json
├── sample-plan.json
├── segments.json
├── transcript.json
├── ocr.json
├── audio-events.json
├── evidence-timeline.json
├── critique.json
└── evaluation.json

The four days

Day	Build	Rust learned by pressure	AI lesson
14	bounded video intake and sampling	files, iterators, checked math	choose evidence before reasoning
15	speech, OCR, cuts, audio events	traits, parallel tasks, typed results	specialist models beat one opaque call
16	aligned segment evidence	interval joins, provenance, serialization	cross-modal reasoning needs coordinates
17	cited critique and expensive-stage routing	policies, budgets, benchmarks	quality/cost is a measured frontier

Run the deterministic planning baseline:

cd rust/rust-ai-engineering
cargo run -q -p mosaic-evidence --example video_pipeline
cargo run -q -p mosaic-ml --example video_generation_funnel

No required step needs a GPU. When a decoder or specialist model is unavailable, fixtures preserve the same typed contract and failure investigations.

Start with Day 14 — Decode and Sample →.