The 34-Day Project Roadmap

This book has two subjects and one rule:

Build the next useful behavior, then learn the Rust and AI concepts that behavior forces into the open.

Project 1 · Brandforge · Days 1–5

Build a business-research and advertising CLI. It accepts a frozen website snapshot and model-produced profile, rejects unsupported claims, compiles an image prompt, and creates a visible preview and run report.

Rust: Cargo, modules, ownership, borrowing, paths, structs, enums, Serde, errors, traits, iterators, files, bytes, hashing, tests.

AI: frozen input, structured output, provider boundaries, evidence grounding, image generation, artifact provenance.

Artifact: a campaign pack you can inspect.

Start: Brandforge overview.

Project 2 · AskDocs · Days 6–9

Build a local RAG CLI that chunks files, embeds them, ranks top-k passages, streams a model response, and verifies citations.

Rust: iterators, slices, lifetimes, float handling, async traits, callbacks, byte buffering.

AI: chunking, embeddings, cosine similarity, retrieval, context manifests, citations, abstention, retrieval and answer evals.

Artifact: cited answers plus a reproducible context manifest.

Start: AskDocs overview.

Project 3 · Signal Room · Days 10–13

Build an incident reconstruction tool from logs, screenshots, traces, deployment changes, and voice notes.

Rust: variant-rich enums, validated constructors, shared bytes, checked integer time, graphs, serialization.

AI: multimodal observations, OCR, transcription, provenance, alignment, hypothesis graphs, uncertainty, adversarial evidence.

Artifact: an evidence-linked incident timeline and hypothesis report.

Start: Signal Room overview.

Project 4 · Cutroom · Days 14–17

Build a video reasoning pipeline with bounded sampling, speech/OCR/audio specialists, aligned segments, cited critique, and expensive-stage routing.

Rust: media clocks, checked math, capability traits, bounded concurrency, interval joins, resource planning.

AI: scene sampling, specialist models, cross-modal alignment, vision-language routing, quality-cost measurement.

Artifact: a segment evidence timeline and verified edit critique.

Start: Cutroom overview.

Project 5 · Forge · Days 18–23

Study the real Rust implementation of OpenAI Codex CLI, then build a safe agent with typed tools, authorization, a bounded loop, effect receipts, causal trace, replay, verified memory, and release evals.

Rust: workspaces, traits, state machines, capabilities, durable events, hashes, policies, async effects.

AI: tool calling, prompt injection, progress, stopping, repair, uncertain effects, memory, trajectory evaluation, release identity.

Artifact: a replayable, eval-gated agent run.

Start: Forge overview.

Project 6 · Model Lab · Days 24–28

Inspect transformer mechanics, run a local inference boundary, route small and strong models, spend test-time compute, and decide whether fine-tuning is justified.

Rust: matrices, numeric validation, tensors, native/runtime boundaries, scheduling, calibrated policy, dataset types.

AI: tokenization, attention, sampling, model artifacts, quantization, local serving, decomposition, routing, generate-check-repair, LoRA/QLoRA and preference data.

Artifact: a measured model route and tuning decision memo.

Start: Model Lab overview.

Project 7 · Mosaic · Days 29–34

Assemble the products into one CLI/API/worker platform with durable jobs, artifacts, streaming, backpressure, tenant isolation, security, observability, rollout, and rollback.

Rust: Axum, SQLx, transactions, leases, channels, semaphores, cancellation, middleware, tracing, containers, graceful shutdown.

AI: durable runs, model/tool budgets, quality-cost-latency signals, safety gates, operational evals, safe degradation.

Artifact: one deployed vertical slice plus failure and recovery evidence.

Start: Mosaic overview.

What “complete” means

Finishing all pages is not the goal. You finish when you can:

build each product from a blank or partial crate;
predict and reproduce its important failure modes;
explain the ownership and async path of expensive data;
replace scripted, local, and remote model adapters without changing domain behavior;
show evidence for factual claims and external effects;
replay agent behavior and evaluate a release;
operate a bounded service and recover it safely.

The Deep Reference Map connects every project pressure to the exhaustive chapters already in the book.