A field guide to discovering customer problems, building integrations, evaluating AI systems, and handing software over to operators. It combines learning material with runnable Python exercises, a local ticket intake reference, and dated source-derived data. Begin with the path that matches the work you need to demonstrate.
Evidence and implementation review: 2026-10-09; beginner curriculum and selected source checks: 2026-10-11. See the audit for confirmed defects, source limits, and the distinction between existing capabilities and production requirements.
- New to programming: start with setup and first run, then follow the six foundations lessons: Python and real JSON, SQL, HTTP, retrieval, and evaluation.
- New to the role: read what an FDE does and use the beginner path or background route selector.
- Already shipping software: use the expert practicum to demonstrate source lineage, holdout discipline, authority boundaries, recovery, economics, and handover.
- Preparing for interviews: start with the interview guide and runnable exercises. Practice rubrics are not authenticated company transcripts.
- Checking market claims: read the market methodology before using counts or salary reports.
- Developing the reference: follow the ETISE README, including its missing production controls.
- FDE glossary - 51 terms with plain definitions and examples covering customer delivery, integrations, AI, evaluation, trust, and operations.
- FDE roadmap - six stages with retained work and exit checks, connecting the beginner lessons, 90-day plan, 24-week plan, and expert practicum.
- Website - open the self-contained learning portal in a browser; lesson and chapter links lead to the repository.
The market snapshot retrieved on 2026-10-09 records immutable upstream revisions, input SHA-256 hashes, source-record references, and nine observation dates through 2026-09-23. Its cumulative title filter matches 212 job IDs across 125 employer names. September has 127 matching listing rows and 91 unique matching IDs. The monthly-file union differs from the cumulative source; the discrepancy is preserved rather than hidden.
The CFPB metadata sample contains five actual complaint records received on 2026-10-09 and retrieved from the official API. It retains categorical metadata without narratives or locations. It is an ingestion and lineage sample, with no invented ETISE quality labels.
The October 9 source ledger retains the earlier primary and publisher checks. The October 11 ledger records selected beginner documentation, pinned inspiration, and an exact official query for complaints received on October 11. That query returned zero hits at capture time, so the teaching sample retains its October 9 date. Both ledgers preserve failed retrievals and claim-specific limits. Verification dates are separate from source publication and collection dates. Neither data collection is a worldwide census, an independently adjudicated customer benchmark, or evidence that every cited vacancy remains open.
The earlier market summary is historical aggregate material, not 146 raw postings. Percentages quoted from that sample cannot be applied to the new sample. ETISE's 25 existing cases are explicitly marked as regression fixtures with unverified origins. Bitext is excluded from real-world-only evidence because its publisher describes it as hybrid synthetic. See the provenance contract.
- Role - responsibilities and title boundaries
- Skills - technical and customer-facing competencies
- Engineering - prototypes, APIs, pipelines, infrastructure, and security
- Customer work - discovery, scope, constraints, and expectations
- AI engineering - application patterns, tools, evaluation, and monitoring
- Deployment - promotion and readiness reviews
- System design - constrained architecture and decisions
- Troubleshooting - diagnosis and incident response
- Case studies - sourced material and clearly identified design exercises
- Job market - dated observations and interpretation limits
- Learning paths - background-specific routes and expert work products
- Interviews - preparation, rubrics, and code
- Portfolio - project selection and evidence presentation
- Resources - further tools, literature, and communities
Existing chapters contain historical claims and recommendations; the audit is not a certification of every external citation. Check evidence status before treating an illustrative scenario, code excerpt, or proposed deployment pattern as an implemented customer outcome.
For a first data exercise, use Python 3.12 from the repository root; this standard-library command needs no package installation:
python3 learning-paths/foundations/code/complaint_pipeline.pyIt reads the five retained real metadata records without creating a database. Continue through the foundations course to write your own function, import into SQLite, and test failures. Local practice outputs belong in the ignored learning-artifacts/ directory.
For the API and full verification, use Linux and Python 3.12 with the tested dependency constraints. From the repository root:
python3 -m venv .venv
source .venv/bin/activate
python -m pip install -r portfolio/reference-project/requirements.txt \
-c portfolio/reference-project/requirements.lock
python -m pip check
python -m pytest interviews/code/ portfolio/reference-project/tests/ job-market/dataset/tests/ learning-paths/foundations/code/ research/tests/ -v
python portfolio/reference-project/evals/run_evals.py --report /tmp/etise-regression.json
python portfolio/reference-project/evals/curate_eval_dataset.py
python interviews/dataset/validate_dataset.py
python interviews/dataset/ten_pass_verification.py
python job-market/dataset/validate_market_data.py
python job-market/dataset/validate_market_snapshot.py
python research/validate_evidence.py
python tools/check_links.pyThe GitHub workflow runs the dependency, test, regression, dataset, and local-link checks on pushes and pull requests, with pinned actions and read-only repository permissions. Its dataset checks are offline accounting checks; they do not refetch or authenticate every external source.
The Python suite tests failure boundaries including payload conflicts, replay scopes, expiry, concurrent requests, exact quotation checks, empty evidence, retry delay, and chunk budgets. The regression report gates routing and required document retrieval as well as classification. It is a known-case regression check, with engine-only latency rather than production SLA measurement.
The interview and historical-market validators check structure and arithmetic. They do not authenticate interviews, company rubrics, source licenses, or a published salary claim. For full market reproduction against the pinned upstream objects, follow the dataset runbook.
ETISE uses deterministic rules, feature hashing, sample policies, and in-memory state. It does not call an LLM, implement BM25, authenticate a tenant, persist an audit log, or dispatch a response to a customer. Its role header is a local simulation. Run it locally with one worker; do not expose it to untrusted traffic.
Production use requires independently labeled customer data, authenticated authority, durable state, resource and retention budgets, operational telemetry, recovery drills, and an accountable handover. The expert practicum makes those gates explicit rather than claiming this repository has already crossed them.
Read STYLING.md. Change a claim only when the supporting artifact or implementation needs to change. Preserve historical observation dates, pin source revisions where available, report disagreements and failed checks, and distinguish observations from recommendations. Do not introduce invented datasets, production outcomes, quotes, or company-specific interview records.
- Foundations - guided first project with expected output and debugging checks
- Expert practicum - experienced-engineer acceptance gates
- Audit - repository findings and remaining production work
- Work in progress - unresolved evidence and implementation gaps
- AI Engineering From Scratch - inspiration for shared foundations and staged practice, inspected 2026-10-11; new lessons are original
- AI Engineering Field Guide - inspiration and source data, retrieved 2026-10-09
- CFPB complaint catalog - original data and interpretation notices, checked 2026-10-09