Portfolio projects that prove FDE skills
Updated
Five builds that map to real FDE work, what each one must contain, and the README habit that reads senior.
A hiring manager spends ninety seconds on your portfolio. These five projects survive that skim because each maps to work FDEs actually do. The portfolio guide sets the bar; this post sequences the builds.
Project 1: the messy-data integrator
Take two inconsistent public datasets, build an idempotent pipeline that reconciles them, and publish the error report format. Must contain: canonical model doc, dedup rules, quarantine design, weekly report template with labeled fictional numbers.
Project 2: the permission-aware assistant
Small retrieval over documents with role-based access, plus the write-up that matters: the leak tests that tried to cross team boundaries and failed. Must contain: access filter inside the query, a 30-question scorecard split by retrieval and generation, adversarial probe results.
Project 3: the flaky-upstream survivor
Wrap an unreliable source: retries with backoff, idempotency keys, dead-letter queue, repair runbook. Must contain: delivery-rate dashboard, duplicate-rate log, one-command replay.
Project 4: the approval-gated workflow
AI drafts, human approves, audit trail records who approved what when. Must contain: action classes, the approval gate demo on fixtures, rollout and rollback plan.
Project 5: the "AI was the wrong tool" case
Rules beat the model on a deterministic task; build rules, compare cost, latency and accuracy honestly. Must contain: the comparison table and the line where you would switch approaches.
Worked example: the README skeleton
Problem in customer terms with one number. Scope: in, out, success metric. Run: three commands. Failure: what breaks and the fix. Next: what production would require. Five headings, one page, no excuses.
Checklist: before you link it
- A stranger ran it from the README alone.
- Errors surface with reasons, never silent drops.
- Numbers are labeled real or fictional.
- One paragraph names what you would harden next.
Related reading
Straight answers
Frequently asked questions
How many projects are enough?
Three end-to-end builds beat eight demos. Each must show brief, code, deployment story and measurement.
Can I use synthetic data?
Yes, and say so. Synthetic or public data stated openly builds trust; hidden customer data destroys it.
What do hiring managers read first?
The README: problem, scope, how to run, what breaks, what you would harden next.