Readiness Self-Assessment
A self-assessment for GenAI engineering roles: score yourself 0-3 on fourteen domains, compare with the target for each, and use the links to the modules and labs that close the gap. Retake it after each module or capstone - the scores are only useful if they are based on evidence you could show someone.
Learning objectives 20 min to score, then use it to plan
By the end of this page you will be able to:- Score yourself 0-3 on each of the 14 domains using the evidence descriptions, not a feeling
- Identify every core domain below 2 and the modules and labs that address it
- Plan three end-to-end projects that show code, metrics, architecture, trade-offs and business outcome
Prerequisites
- None - take it before starting the course and again after each track
The Scale
| Score | Meaning |
|---|---|
| 0 | Cannot explain or build it |
| 1 | Recognize the concepts; can follow a tutorial |
| 2 | Can build it independently |
| 3 | Can design it, debug it and explain the trade-offs |
Readiness threshold: no score below 2 in core areas, and at least three end-to-end projects where you can show the code, metrics, architecture, trade-offs and business outcome.
Domains, Targets and Where to Learn Them
Coverage says how fully this course covers the domain: Full means the notes and labs can take you to the target. Every domain is now covered in full; the links point to the notes and labs that build each one.
What a 2 and a 3 Look Like
Score yourself on what you have actually done, and could show.
| # | You're at 2 when you can... | You're at 3 when you can also... |
|---|---|---|
| 1 | Build a typed, tested Python service with a REST or streaming API, containerize it, and work on a remote Linux GPU machine with Git branches and PRs | Design the API contract (streaming, idempotency, rate limits, versioning), and debug a container, driver or dependency problem from first principles |
| 2 | Deploy a model server on Kubernetes with Helm, GPU scheduling and least-privilege IAM | Design a private, multi-environment deployment (networking, identity, IaC, autoscaling signal) and defend it against alternatives on another cloud |
| 3 | Work through attention shapes, softmax and cross-entropy by hand, and compute a confidence interval for an eval score | Explain optimizer behaviour, low-rank adaptation and training instabilities mathematically, and choose the right statistical test for a comparison |
| 4 | Write a training loop from scratch and fine-tune a Hugging Face model with your own data | Diagnose a training run that won't converge, or a memory blow-up, from the code and the curves |
| 5 | Implement a small GPT and explain each component, and how a tokenizer turns text into ids | Explain why an architecture choice (GQA, MoE, RoPE scaling, vocabulary size) changes quality, memory or cost, with numbers |
| 6 | Run LoRA/QLoRA and preference tuning on curated data, and show a measured improvement over the base model | Decide between prompting, RAG, SFT, DPO and RL for a problem, and design the data pipeline and checks for it |
| 7 | Build a task-specific eval set, use a validated LLM judge, and run a red-team scan reporting ASR and over-refusal | Design an evaluation and safety programme for a product - gates, online evals, attacker budgets, grader validation - and explain what each number can't tell you |
| 8 | Estimate the GPU memory of a training or serving job and fix an out-of-memory error | Profile a slow step, identify whether it is compute-, memory- or communication-bound, and fix it |
| 9 | Run a multi-GPU training job with DDP or FSDP | Choose a parallelism strategy (data, tensor, pipeline, ZeRO stage) for a model and cluster, and estimate the communication cost |
| 10 | Serve a model with vLLM behind an API and load-test it | Tune batching, KV cache, precision and parallelism against an SLO, and explain the goodput trade-offs |
| 11 | Build and evaluate a RAG pipeline with hybrid retrieval and reranking | Design retrieval over enterprise sources with permissions, freshness and governance, and measure and fix retrieval failures |
| 12 | Build a tool-using agent with an MCP server, bounds and human approval | Design a production agent's security model (injection, trifecta, credentials), durability and evaluation, and choose workflow vs agent with evidence |
| 13 | Instrument a service with traces and metrics and set an SLO with alerts | Run an incident and a blameless postmortem, design burn-rate alerting and cost controls, and catch silent quality regressions |
| 14 | Run a discovery conversation and write a design doc with diagrams and ADRs | Qualify a use case with an ROI model, present it to executives and engineers, and plan a pilot with exit criteria to production |
Three End-to-End Projects
The Capstones are designed as the three projects - each asks for code, metrics with uncertainty, an architecture document with ADRs, trade-offs and a business outcome.
| Capstone | Domains it gives evidence for |
|---|---|
| 1 - Train and Post-Train a Small Model | 3, 4, 5, 6, 7, 8, 9 |
| 2 - Serve It with an SLO | 1, 2, 8, 10, 13 |
| 3 - Ship a Production Agent | 7, 11, 12, 13 |
| All three | 14 - every capstone requires a design document, ADRs and a business outcome |
If a domain you need is not covered by these, build a fourth project around it - for example, a permission-aware RAG system over a real document set for domain 11.
How to Use It
- Score all fourteen domains honestly, using the evidence columns above.
- List every core domain below 2. Those come first, before polishing domains already at 2.
- For each, work through the linked notes, then do the lab - reading moves you from 0 to 1; building moves you to 2.
- Move to 3 by explaining trade-offs to someone else: the module Q&A banks and the system designs are practice for that.
- Rescore after each track, and keep the evidence (repositories, reports, dashboards) - it is what you will show in interviews.