Portfolio case study & pitch

What you'll produce

A one-page portfolio case study that compresses your five prior Support Copilot deliverables — prompt spec, eval, risk register, PRD, and the RAG build — into a single, recruiter-readable narrative, plus a 90-second spoken pitch script you can deliver from memory in an interview. The case study is the artifact a hiring manager actually opens first: it's the link on your resume and the thing they skim in 60 seconds before deciding whether you're worth a call. For an AI-adjacent candidate this is the highest-leverage page you own, because it proves the exact thing Topic 10 says employers most want from a non-traditional hire — that you can put AI in front of real users and know where it must not act alone. Done well, it tells one coherent story (problem → built → evaluated → de-risked → impact) and lands the domain-expertise + AI-judgment frame that separates you from someone who has only watched prompt-engineering videos. The pitch script turns that same story into something you can say out loud, calmly, in the first five minutes of a screen.

Instructions

1. Gather your five artifacts and pull the load-bearing numbers. Open Deliverables 1–5 and extract the specifics that make the story credible: the baseline (14-hour reply time, doubled ticket volume), your eval scores (the pre-fix pass rate — 1 of 7 drafts in D2 — and the dimensions that scored worst), your top failure pattern (e.g. "invents fee amounts"), the guardrail that kills that pattern, your "good enough" quality bar from the PRD (D4's ≥95%-accuracy release gate), and the result of the RAG build. Use the actual numbers from your deliverables — don't round, soften, or invent a cleaner-sounding figure. The fastest way to lose a sharp hiring manager is a headline number they can't find when they click into the work; a case study whose figures trace back to the source beats one with adjectives every time.
2. Write a one-line hook (the headline). State the feature and the tension in a single sentence a busy reader can grasp instantly — name the domain (fintech support) and the central risk (wrong answers about money). This is the first thing the hiring manager reads; make it earn the next ten seconds.
3. Write the Problem section (2–4 sentences). Quantify the pain (volume, reply time, the cost of getting a money answer wrong) and state the vague brief you were handed ("use AI to fix this"). Establish that the hard part was judgment, not typing a prompt.
4. Write the What I built section. Describe the Support Copilot in one or two sentences, then list the five artifacts as a build arc — one line each, each naming the skill it proves. This is where you show range: prompt engineering → evaluation → risk → spec → build.
5. Write the How I evaluated and de-risked it section. This is the section that separates you from prompt-tutorial graduates. Name your rubric dimensions, give the headline score honestly (including what failed), name the #1 failure pattern, and state the concrete guardrail + human-in-the-loop decision that contains it. Honesty about a real failure is a hiring signal, not a weakness — say what broke.
6. Write the Impact section. Tie it to a measurable outcome and the quality bar — and pull both straight from your deliverables rather than inventing a round number. Your honest pre-fix evidence is D2's 1-of-7 pass rate (14%), and your quality bar is D4's ≥95%-accuracy release gate (with D2's re-run target of ≥90% pass and zero Safety/Factuality failures). If the build is a prototype, frame the impact as the shape of the change the eval proves — "the grounding-plus-guardrail fix moves the dangerous failures from caught-in-review to caught-by-the-system, so agents review-and-send instead of write-from-scratch" — rather than claiming a production acceptance percentage you don't have. A specific, defensible "before" number you can stand behind beats an impressive "after" number a reader can't verify.
7. Write the Domain + AI framing section (your differentiator). One short paragraph in your own voice: why your background + AI literacy makes you the right person to own a feature like this. This is the Topic 10 thesis made personal.
8. Add a short links/artifacts list so a reader can click into the underlying work (the five deliverables), and a one-line tools/concepts tag (LLM, RAG, eval rubric, human-in-the-loop) so the keywords a recruiter scans for are present.
9. Write the 90-second pitch script (~210–250 words spoken). Structure it: hook (who you are + the project) → problem → what you did → the one impressive specific (your best failure-catch + fix) → impact → why-you. Write it to be said, not read — short sentences, no jargon you can't define if interrupted. Use the same numbers as your written case study, said in words ("one of seven drafts," "ninety-five percent accuracy") — a spoken figure that contradicts your page is the fastest way to look like you don't know your own work. Time yourself reading it aloud; trim to land under 90 seconds at a calm pace.
10. Pressure-test for coherence. Re-read and confirm the thread is visible: the eval's failure pattern drives the risk register's mitigation, which drives the PRD's acceptance criteria. If a hiring manager can't trace that line, tighten it — coherence across the six deliverables is exactly what graders and employers reward.

Rubric

The app's AI scores the learner's submission against these criteria and gives feedback. Levels: Needs work (1) / Solid (2) / Excellent (3). Passing = every criterion at Solid or above.

• Coherent narrative arc with a traceable thread — 1: a list of artifacts with no through-line, or figures that contradict the underlying work · 2: follows problem → built → evaluated → impact · 3: one tight story where the same failure (the D2 instant-transfer draft) and the same numbers (D2's pass rate, D4's quality bar) carry through the eval, the risk mitigation, the PRD criterion, and the rebuild — a reader who clicks into any of the five prior deliverables finds the figure exactly as cited.
• Concrete, honest evidence that traces to the source — 1: vague adjectives ("worked well," "made it safe"), or numbers that don't appear in any prior deliverable · 2: real numbers from the deliverables (14-hour reply time, the D2 pass rate, the D4 quality bar) · 3: load-bearing specifics including a named failure tied to an actual eval draft (e.g. the invented "$25" flat fee on the instant-transfer ticket from D2 Draft 1, where the true capped fee was $15) and the exact guardrail that contains it — every figure (pass rate, quality bar) reconciles with the deliverable it cites, and prototype-vs-production claims are stated honestly.
• De-risking & human-in-the-loop is front and center — 1: risk/HITL absent or an afterthought · 2: mentions a mitigation and where a human reviews · 3: a specific guardrail ("only state a number present in the retrieved snippet; otherwise refuse and escalate") plus a clear no-auto-send / mandatory-review decision and the AI-disclosure policy — the signal these roles hire for.
• Domain + AI framing — 1: generic "I'm passionate about AI" · 2: names the domain and pairs it with AI literacy · 3: makes the Topic 10 thesis personal and credible — why this specific background + hands-on AI skill makes them the right owner of a money-sensitive AI feature.
• 90-second pitch is deliverable out loud — 1: a written essay that can't be spoken in 90s, or jargon-stuffed · 2: hits hook → problem → what I did → impact in roughly the right length · 3: tight spoken English (short sentences, a memorable failure-and-fix beat), genuinely lands under 90 seconds, and ends on a why-you line a hiring manager remembers.
• Recruiter-readable & self-contained — 1: needs the reader to already know the project · 2: one page, scannable, links to the artifacts · 3: a hiring manager grasps the whole story in 60 seconds, the right keywords (LLM, RAG, eval, human-in-the-loop) are present for screening, and every artifact is linked for a deeper look.