Become an AI Expertv2.1
0 of 0 steps Continue
Reference

Project ladder

Every project in the course, in order, with what it teaches, a time estimate, and the "done when" bar. Both tracks climb the same ladder; the Module column says where each rung is built, with the Novice page in parentheses. Project numbers are the ones the modules use.

The ladder follows the practitioners' order: a tiny model from scratch (Raschka: "implementing a simple model from scratch that you can run on your computer"), then reproduce end-to-end projects yourself (Boykis), then ship something a friend uses (Huyen), then an LLM app where you do error analysis by hand (Husain), then projects that "break in interesting ways" with evals, tracing and deployment.

# Project Module Teaches Time Done when
1 Deployed Python service with tests and Docker 01 (A1) Software fundamentals, deployment, reading tracebacks 20 to 35h A stranger can clone, test and hit the live URL
2 Gradient descent from memory in NumPy 02 (A2) What training is 4 to 6h Both fits converge; you can draw it on paper
3 Tabular project with cross-validation and error analysis 02 (A3) Framing, leakage, metrics, looking at your data 15 to 25h You can say with numbers why it beats the baseline and where it fails
4a Your own tokenizer and n-gram model 03 (A4) Tokens, the next-token objective 6 to 10h Lossless round trip; the vocabulary trade-off table
4b Your GPT, retyped from memory, trained and sampling 04 (A4) The Transformer 15 to 20h It trains and samples; the forward pass whiteboarded with shapes
5a nanochat run with eval table and one measured change 05 (A4) Pretraining, scaling laws, cost 20 to 40h plus $20 to $75 compute Eval numbers, loss curve, cost line, written up with what broke
6 KV cache in your GPT with speedup numbers 05 (A4) Inference, memory 6 to 10h Identical outputs proved; before/after timing in the README
5b SFT and GRPO on your model with separate eval columns 06 (A4) Post-training 15 to 25h plus compute Capability and general-ability columns per stage, and what the KL coefficient did
7 One ablation (one variable, three settings, one plot) 06 (A4; Track C) Experimental method 6 to 10h A plot and a paragraph someone else can check
8 Retrieval over a messy corpus with a 50-question golden set 07 (A5) Retrieval, evals, permissions, cost 30 to 50h recall@k before/after one change; five real users for two weeks
9 The harness (part 1), the agent (part 2), the red-team suite (part 3) 08, 09, 10 (A5) Evals, agents, context engineering, security 40 to 60h pass@3 and pass^3 reported; unfixed failures documented; the suite in CI
10 A cost cut with numbers 10 (A5) Caching, routing, token budgets 8 to 15h Cost per request halved without hurting the eval
11 Fine-tune versus not: three variants and a decision record 11 (A5) When fine-tuning wins 10 to 15h plus under $20 Quality, cost and latency for each variant; the decision follows the numbers
12 The pilot: pitch, delivery, results memo 12 (A6) The internal move 20 to 40h Eval numbers, cost and failure cases in the memo
13 The public record: three posts, a finished README, a talk, one PR opened 12 (A6) Being known 15 to 25h A stranger has starred, cited or commented
14 The application kit: portfolio page, numbered resume, timed design prompts 13 (A6) Getting hired 10 to 15h Three referral conversations, one interview
15 Track artifact (A: agent in production with a dashboard; B: Triton kernel with a fair speedup; C: replication plus ablation) plus one merged upstream PR 14, 15 or 16 (A7) Depth 24h guided, then months As the track module defines it; the PR is merged
16 The yearly plan 17 (A8) Staying current 2 to 4h Names an artifact and a date a stranger can check

Project ideas that are not "chat with your PDF"

From the research, projects that practitioners say teach something and "break in interesting ways":

  • Document Q&A with citations and an eval set over a corpus people actually ask questions about (a club's rules, a product's manuals, a city's council minutes)
  • A support-ticket triage workflow with tools and structured outputs, graded on routing accuracy
  • A receipt or invoice parser with validation and a measured error rate per field
  • A small agent that plans, acts, checks and retries within a token and dollar budget, with the budget enforced in code
  • Churn or fraud prediction on real business-shaped data, which forces class imbalance and a cost-weighted metric
  • Anything with time in it (stock prices, sensor data), which forces you to learn temporal leakage the hard way
  • Something from your own frustration log that a friend will use

What makes a project read as finished

Reviewers of 2026 portfolios converge on this. A project with none of it "reads as unfinished."

  • One-line result with a number at the top of the README
  • An eval table against a baseline, on a described and versioned test set
  • p50 and p95 latency and cost per request
  • A "where this breaks" section with real failure cases
  • Decisions you reversed and the measurements that drove them
  • A live URL or a one-command run. An un-runnable project reads as broken.
  • Clear statement of what is yours if it builds on something else. Do not post course exercises as projects.

Sources

Free to use: code MIT, writing CC BY 4.0, third-party media under its owners’ terms. Licence and terms.