AI Engineer Notebooks – free, framework-free RAG/agents/evals on Colab

Learn the applied-LLM stack the measurement you'll really beryllium interviewed connected it — framework-free, connected a free API, from prompting each the measurement to serving, fine-tuning, and a red-team benchmark.

Runnable Colab notebooks for the AI Engineer / Forward Deployed Engineer (FDE) skill set: building moving systems connected apical of instauration models — exemplary APIs, RAG, evals, agents, adaptation, serving — utilizing earthy APIs, not frameworks.

• Framework-free, connected purpose. You constitute the supplier loop, RAG, and evals from raw API calls first — truthful you understand what LangChain/LlamaIndex really do before you scope for them (and tin judge erstwhile not to). Patterns are durable; wrappers churn.
• Evals are the spine. "Measure earlier you tune" is installed early and returns successful each conception — the wont that separates an technologist who shipped a system from 1 who built a demo.
• Free to run, extremity to end. Everything runs connected the free Groq API (no in installments card). The 2 topics Groq can't big — LoRA fine-tuning (06) and self-hosted serving (09) — are concept-first pinch optional, fenced Colab-GPU appendices, verified connected a real Colab T4.
• Real lawsuit studies, not artifact demos. Three end-to-end case studies show the skills mixed under real constraints — a support adjunct debugged successful production, a pipeline-vs-agent costs showdown, and a red-team robustness benchmark.
• OpenAI-compatible throughout, truthful each shape transfers straight to OpenAI and (with mini changes) Anthropic — the seam is swappable, the skills aren't.

Built arsenic the hands-on companion to Plan: Transitioning to Forward Deployed Engineer / AI Engineer. The scheme explains what to study and why; these notebooks are wherever you tally it.

Backend aliases full-stack engineers moving into AI Engineer, FDE, Applied AI, or Solutions Engineer (AI) roles — different titles, mostly the aforesaid job. You tin vessel accumulation code; you want the applied-model furniture connected top.

Work apical to bottom. Each notebook is self-contained (installs its own dependencies, sounds API keys from Colab secrets) and ends pinch exercises.

Notebook What you'll learn

Measuring outputs	Golden sets and metrics connected the section-01 task — instal the "measure earlier you tune" wont before building thing you'd request to tune. Evals is the spine; it returns successful each conception aft this

Notebook What you'll learn

Fine-tune vs RAG vs prompt	When to alteration the model's weights vs its inputs; what LoRA/QLoRA are and cost; the statement you'll person successful the room — positive an optional existent LoRA fine-tune connected a free GPU

Notebook What you'll learn

Prompt injection & the spot boundary	Direct & indirect punctual injection, output handling, PII, excessive agency — the OWASP LLM Top 10 risks, failing unrecorded past defended

Notebook What you'll learn

Observability & LLMOps	Tracing each call, safe punctual logging, cost/latency/error metrics, drift detection, and the observe→eval feedback loop
Reliability & fallbacks	Retries pinch backoff, timeouts, fallback models, output validation, circuit breakers, graceful degradation
Experiment search & registry	MLflow extremity to end: log runs/params/metrics from the section-04 eval harness, registry and type a model, and beforehand by shape — the tooling that turns "I ran an eval" into a tracked, reproducible workflow

Where the free Groq API can't tally the taxable (these frameworks request a GPU), the notebook teaches it concept-first and fences an optional Colab-GPU appendix — the aforesaid shape arsenic the section-06 LoRA appendix.

Notebook What you'll learn

Designing an conclusion service	Concept: the ML strategy creation interview, worked extremity to extremity — QPS/VRAM/latency/cost estimation, replica scaling, queueing, caching, and the SLA trade-offs, connected a realistic LLM-serving prompt

Notebook What you'll learn

Scoping & discovery	Turn a vague customer inquire into a scoped, evaluable system: find questions, a one-page scoping doc, the demo subject — the customer-scenario question and reply information astir engineers can't evidence

Where the skills travel together into projects. First a case study — one realistic script worked extremity to end, runnable — past the capstone, the deployed repo you build yourself. (Section overview.)

Capstone: the brief for the deployed project that goes connected your resume — a existent repo pinch a serving constituent and an eval report. Case studies are for learning; the capstone is for hiring.

• Raw exemplary APIs, nary frameworks. Patterns are durable; wrappers churn.
• One shared corpus (data/) crossed RAG and eval sections, truthful evals measure the retrieval you really built.
• Self-contained notebooks. First compartment installs, 2nd compartment calls from aien import setup; client, MODEL = setup() to load your cardinal from Colab secrets (or a section env var). No hidden authorities betwixt notebooks. aien is the mini shared-setup package successful this repo — 1 spot to change credential loading — installed automatically by the first cell.
• Every notebook ends pinch exercises — do them earlier moving on.

1. Get a free API cardinal astatine console.groq.com — no credit paper required.
2. In Colab: the cardinal icon successful the near sidebar → adhd GROQ_API_KEY arsenic a secret, and toggle notebook entree on.
3. Open immoderate notebook via its badge and tally apical to bottom.

Running locally instead: pip instal -r requirements.txt && pip instal -e . (the 2nd installs the aien setup helper), export GROQ_API_KEY=..., open pinch Jupyter.

• Plan: Transitioning to FDE / AI Engineer — the roadmap these notebooks implement
• Guide: Building a Real LLM Project for Your Resume — the capstone's requirements bar
• Walkthrough: Designing a RAG System — the systems position of conception 03
• Walkthrough: Designing an AI Agent Orchestration System — the systems position of conception 05