Engineering Notebook

Senior-engineer review surface

Production-shaped AI prototypes, grounded labs, and a visible quality gate.

This workspace demonstrates practical AI engineering through constrained agents, runnable RAG, developer-experience tooling, data pipelines, repeatable labs, and automation that keeps the artifacts honest.

Open README Inspect Agent Traces
workspace quality gate
$ python scripts/validate_workspace.py
enablement-assistant tests       PASS
enablement-assistant eval        18/18
agentic-workflow tests           PASS
agentic-workflow eval            7/7
prompt-regression tests          PASS
prompt-regression comparison     PASS
local-ai-lab-runner tests        PASS
lab readiness checks             PASS

All workspace validation checks passed.
8 Production-shaped projects
5 Standardized labs
1 Workspace validation command
0 Required external services for baseline demos

Runnable projects

Evidence over claims

Live on Vercel Verified workflow

Support Triage Review Console

Turns a synthetic support ticket into a validated model suggestion and an explicit human-approved decision with visible operational evidence.

  • Access-code live path limited to three synthetic cases
  • Shared prompt and exact schema from the regression suite
  • Open recorded mode plus a shared daily spend ceiling
  • End-to-end provider success and safe failure diagnostics verified
  • Browser-local accept/correct feedback with sanitized JSON export
Open live studio Recorded mirror Inspect evaluation Inspect feedback loop Deploy your own
Runnable Customer Integration

Mocked Zendesk-Style Support Adapter

Receives signed synthetic support webhooks, blocks replay, redacts PII, maps tickets into the triage workflow, and stops at a proposed update.

  • Strict webhook and ticket contracts
  • HMAC verification and duplicate-event protection
  • No account, external service, or ticket mutation
Inspect integration traces Read project record
Runnable Prompt Evaluation

Prompt Regression Runner

Compares vague and structured prompts across the same support tickets, exposing schema, classification, grounding, and actionability regressions.

  • 15 fixed cases and eight checks per case
  • Credential-free recorded baseline plus optional live adapter
  • Facilitator guide for customer-facing enablement
Open comparison Open feedback candidates Open trainer guide
Runnable Agents

Agentic Workflow Demo

Uses narrow backlog tools, pauses before simulated mutations, refuses prohibited actions, and exposes every decision in a structured trace.

  • Application-owned tool permissions
  • Human approval checkpoint with defense in depth
  • Behavioral evals for tools, sources, and refusals
Open trace viewer Read project record
Runnable RAG

Enablement Assistant RAG

Answers from trusted markdown sources, returns citations, exposes retrieved context, and refuses unsupported questions.

  • Source-aware chunking and retrieval
  • Grounded answer contract
  • Evaluation set for retrieval and refusal behavior
Open demo View eval report Read project record
Runnable Developer Experience

Local AI Lab Runner

Checks local readiness, lab structure, runtime state, package imports, credentials, and checkpoint guidance before experiments begin.

  • PASS/WARN/FAIL findings
  • JSON output for automation
  • Simulated failure modes for demos
Read project record
Scaffolded Data Engineering

StreamFlow Phase 1

Containerized stream processing platform with a synthetic producer, Redpanda, Spark Structured Streaming, Airflow, and data quality rules.

  • Kafka-compatible event flow
  • Checkpointed Spark ingest and batch summary
  • Rejected records with explicit reasons
Open report
Scaffolded Analytics Engineering

StreamFlow Phase 2

Enterprise analytics pipeline with Snowflake Bronze, Silver, and Gold layers, Power BI measures, Airflow orchestration, and reconciliation checks.

  • Medallion architecture SQL
  • Gold star schema for reporting
  • Dashboard QA and DAX contracts
Open dashboard Open report

Labs

Repeatable learning loops

Each lab follows the same readiness contract: objectives, prerequisites, setup, exercise, expected output, failure modes, build notes, and debrief.

Quality gate

What senior reviewers can trust quickly

Single command

python scripts/validate_workspace.py

Runs link checks, eval freshness, project tests, RAG evals, lab readiness, and an optional engineering-quality gate.

CI ready

.github/workflows/validate.yml

Runs pinned Ruff, mypy, 80% branch-aware coverage, seven isolated Chromium workflow tests, and the complete workspace validation on push and pull request.

Operational notes

04-explainers/debugging-playbook.md

Shows how environment, credential, retrieval, and agent failures are triaged.

Open playbook

Evaluation and observability

01-concepts/evaluation-observability.md

Separates deterministic tests, behavioral evals, request traces, and production monitoring.

Read concept guide

System review checklist

05-reference/ai-system-review-checklist.md

Turns workflow fit, data, oversight, evaluation, security, reliability, and cost into review questions.

Use checklist

Deployable surface

Static by design

This app can be served by GitHub Pages with no runtime server, no secrets, and no package install. The workflow uploads the `docs/` directory as the Pages artifact and deploys it from GitHub Actions.

01 Commit static app 02 Enable Pages from Actions 03 Run Deploy Showcase workflow