Overview — Project 1: Brandforge

The fastest way to understand Rust and AI together is to make something you can inspect. In this project you will build Brandforge, a command-line product inspired by a small TypeScript experiment: give it business research, ask a model for a brand profile, reject unsupported advertising claims, and create a campaign pack.

The model is allowed to propose. Rust decides what is valid, what is supported by evidence, and what may become an artifact.

What Brandforge does

Run the finished reference crate:

cd rust/rust-ai-engineering
cargo run -p brandforge -- build \
  --source fixtures/brandforge/site.md \
  --profile fixtures/brandforge/business-profile.json \
  --output target/brandforge

It writes:

target/brandforge/
├── business-profile.json   # typed model output
├── claim-audit.json        # evidence result for every proposed claim
├── image-prompt.txt        # compiled only from approved information
├── ad-preview.svg          # visible preview you can open in a browser
└── run-report.json         # input hashes and artifact inventory

Now run the dangerous case:

cargo run -p brandforge -- build \
  --source fixtures/brandforge/site.md \
  --profile fixtures/brandforge/business-profile-unsupported.json \
  --output target/brandforge-bad

The model profile claims 98% client satisfaction, but the website snapshot never says that. The program exits unsuccessfully, preserves claim-audit.json, and refuses to create an ad preview. That failure is part of the product.

The five days

Day	Product slice	Rust that the slice forces	AI lesson
1	CLI plus website snapshot	Cargo, modules, PathBuf, String, borrowing	freeze model input
2	typed business profile	structs, enums, Serde, Option, validation	structured output is still untrusted
3	replaceable profile producer	traits, errors, configuration	scripted first, live provider second
4	evidence gate	iterators, normalization, exhaustive state	prompts do not verify facts
5	campaign pack and preview	bytes, files, hashing, unit and CLI tests	artifact provenance and evals

The pipeline you will own

website snapshot ───────┐
                       ├─▶ claim audit ─▶ verified prompt ─▶ preview / image
model business profile ┘        │
                                └─▶ reject unsupported campaign

There are two important boundaries:

1. The model output is decoded into BusinessProfile, but decoding only proves its shape.
2. Every promotional claim must point to a quote present in the frozen source before generation.

The live-provider extension may use an API, but the required build is offline. You can learn the software, rerun the tests, and diagnose failures without a key, network, GPU, or paid request.

Code layout

crates/brandforge/
├── src/
│   ├── main.rs       # CLI and exit status
│   ├── lib.rs        # public product boundary
│   ├── profile.rs    # model-facing data and claim audit
│   ├── pipeline.rs   # read, validate, compile, write
│   ├── render.rs     # image prompt and SVG preview
│   └── error.rs      # named failure modes
└── tests/cli.rs      # run the real executable

The reference crate is the finished answer. For the strongest learning loop, create a scratch binary and type each day’s slice before comparing it with the reference.

Start with Day 1 — CLI and Owned Input →.