AI Engineering Notebook

A personal lab for building modern AI software, writing down what works, and turning experiments into small demos.

This is not a course repo or a role-prep packet. It is a builder's notebook: concepts, prototypes, debugging notes, patterns, and project writeups that show practical fluency with LLMs, RAG, agents, APIs, SDKs, cloud workflows, and AI product design.

Start Here

1. Open the deployed Support Triage Review Console: https://ai-engineering-notebook.vercel.app/support-triage. Use recorded mode for the public, zero-cost walkthrough; the access-code path makes a real model call, and review actions capture human feedback without another call.
2. Inspect the Agentic Workflow trace viewer: https://pablodcruz.github.io/ai-engineering-notebook/docs/agentic-workflow.html.
3. Use the GitHub Pages Support Triage mirror when you want the explicitly labeled recorded experience with no live backend.
4. Compare prompt candidates in the Prompt Regression report: https://pablodcruz.github.io/ai-engineering-notebook/docs/prompt-regression-report.html.
5. Follow human corrections into candidate eval cases: https://pablodcruz.github.io/ai-engineering-notebook/docs/feedback-candidate-report.html.
6. Inspect the mocked customer-system integration: https://pablodcruz.github.io/ai-engineering-notebook/docs/support-adapter.html.
7. Try the Enablement Assistant demo: https://pablodcruz.github.io/ai-engineering-notebook/docs/enablement-assistant.html.
8. Review the Enablement Assistant eval report: https://pablodcruz.github.io/ai-engineering-notebook/docs/enablement-eval-report.html.
9. Open the StreamFlow analytics dashboard: https://pablodcruz.github.io/ai-engineering-notebook/docs/streamflow-dashboard.html.
10. Use 00-meta/roadmap.md to pick the next experiment.
11. Capture technical understanding in 01-concepts.
12. Build small runnable exercises in 02-labs.
13. Promote the best ideas into 03-projects.
14. Turn useful explanations and diagrams into 04-explainers.
15. Collect polished demos and writeups in 06-showcase.

For a systematic architecture review, use the AI System Review Checklist. For the distinction between tests, evals, traces, and monitoring, read Evaluation And Observability.

Fast Validation

Run the full local quality gate:

python scripts/validate_workspace.py

This checks documentation links, eval export freshness, runnable project tests, the Enablement Assistant evaluation set, and every lab's readiness contract.

Run the complete CI-equivalent engineering gate:

python -m pip install -r requirements-dev.txt
python scripts/validate_workspace.py --quality

This additionally enforces Ruff linting and formatting, practical mypy coverage over core modules, and an 80% branch-aware unit-test coverage floor. See QUALITY.md for scope and exclusions.

Run the isolated recorded-mode browser suite:

npm ci
npx playwright install chromium
npm run test:browser

The Playwright suite uses its own Chromium build and no application secrets or billable model calls. See BROWSER_TESTING.md.

Showcase App

Open the canonical live application:

https://ai-engineering-notebook.vercel.app/support-triage

The live path is deliberately bounded: it accepts three synthetic cases, requires an access code, validates an exact response schema, uses a shared Redis daily ceiling, and exposes safe request telemetry. Recorded mode is always available without a provider call.

View the static portfolio and zero-cost mirrors:

https://pablodcruz.github.io/ai-engineering-notebook/docs/

View the StreamFlow analytics dashboard:

https://pablodcruz.github.io/ai-engineering-notebook/docs/streamflow-dashboard.html

Try the Enablement Assistant RAG demo:

https://pablodcruz.github.io/ai-engineering-notebook/docs/enablement-assistant.html

Review the Enablement Assistant eval report:

https://pablodcruz.github.io/ai-engineering-notebook/docs/enablement-eval-report.html

Source lives in docs/index.html.

Vercel is the canonical runtime for the Support Triage Review Console because it serves the Python API functions. GitHub Pages serves the static portfolio and explicitly recorded demos under /docs/. See docs/DEPLOYMENT.md.

Senior Engineer Review Path

If you are reviewing this workspace quickly, start with:

1. Senior Review Guide for the review rubric and strongest evidence path.
2. Agentic Workflow Demo and trace viewer for tool contracts, approval boundaries, refusal behavior, and observable execution.
3. Prompt Regression Runner, comparison report, and facilitator guide for prompt versioning, structured regression checks, and trainer delivery.
4. Support Triage Review Console, feedback candidate report, self-deployment guide, and operating-economics explainer for a customer-facing workflow that connects generation, validation, human correction, reviewed evaluation candidates, cost controls, and operations.
5. Mocked Zendesk-Style Support Adapter and integration trace viewer for signed webhooks, strict mapping, replay protection, PII handling, and proposed-only customer-system updates.
6. Enablement Assistant RAG, deployed RAG demo, and eval report for grounded-answer architecture, evaluation, and citations.
7. Local AI Lab Runner for developer experience, readiness checks, and CLI design.
8. StreamFlow Phase 1 for containerized stream processing, Spark, Airflow, and data quality.
9. StreamFlow Phase 2 for Snowflake medallion layers, analytics modeling, Power BI semantics, and reconciliation checks.
10. Showcase Index for the strongest demo-ready artifacts.
11. Debugging Playbook for operational thinking and troubleshooting judgment.

Workspace Map

Folder	Purpose
00-meta	Roadmap, build themes, and notebook conventions.
01-concepts	Notes on LLMs, prompting, RAG, agents, APIs, cloud workflows, evaluation, and observability.
02-labs	Small experiments with setup steps, expected output, and failure notes.
03-projects	Larger builds that combine concepts into useful AI tools.
04-explainers	Clear writeups, diagrams, and talk tracks for explaining what was built.
05-reference	Glossary, resource log, and a reusable AI-system review checklist.
06-showcase	Polished project summaries, demos, screenshots, and portfolio-style notes.

Build Loop

Each cycle should produce something concrete:

• Learn one concept deeply enough to explain it.
• Build one small thing that exercises the concept.
• Break it on purpose and document the failure mode.
• Improve the design, prompt, retrieval flow, tool boundary, or developer experience.
• Write down the practical takeaway.

Current Themes

• Prompt design for reliable structured outputs.
• Retrieval-augmented generation over small trusted corpora.
• Tool-using agents with narrow permissions and observable steps.
• API and SDK integration patterns.
• Cloud-adjacent AI app architecture.
• Data engineering and streaming platform foundations.
• Analytics engineering with warehouse modeling and BI validation.
• Debuggable developer workflows for AI prototypes.