# Interview stories (STAR-ready)

Use these as spoken stories, not scripts. Each has a punchline a hiring manager can remember.

---

## 1. FlyingMinds architecture

**Situation.** FlyingMinds needed an always-on Learning Coach (Fred) for K–10 — diagnose, adapt, advance — but the early stack couldn’t represent mastery honestly or stay affordable at $29/month.

**Task.** As sole engineer, rebuild the tutoring engine so practice routing, explanations, grading, payments, and admin could run as one product.

**Actions.**
- Modeled skill mastery with **Bayesian Knowledge Tracing** so every answer updated state, not vibes.
- Put a thin **Node/Express** API in front of session logic, Postgres/Supabase persistence, and a DeepSeek LLM layer for explain/grade/generate.
- Shipped **Stripe + admin** early so the business loop existed while curriculum expanded.
- Kept curriculum chrome behind revenue signal — tutor loop first.

**Result.** Scaled to **1,500+ students** and **$15K+** revenue without a second engineer. Architecture interview hook: *mastery store → router → LLM → eval feedback*, not “we wrapped ChatGPT.”

**If they dig:** tradeoffs on DeepSeek vs stronger/more expensive models; why BKT over simpler counters; what you’d split into services at 10× scale.

---

## 2. LLM evaluation (grading agreement 81% → 92%)

**Situation.** LLM grading disagreed with human keys ~19% of the time. Silent errors would corrupt BKT mastery and parent trust.

**Task.** Make grading reliable enough to drive adaptive routing in production.

**Actions.**
- Built a **labeled eval harness** (fixed items, human keys, automated runs).
- Measured **agreement**, error modes (false credit vs false fail), and regressions per prompt/rubric change.
- Iterated rubrics and prompts against the harness — not against vibes from a few chat transcripts.
- Only promoted changes that moved the metric without tanking coverage.

**Result.** Key agreement **81% → 92%**. Cultural change: evals before prompt poetry.

**Demo for interviews:** walk the `showcase/llm-eval-harness` repo — same idea, public and small.

---

## 3. MaxedS automation

**Situation.** Fractional sales team lived in a clunky CRM and burned hours on document triage and handoffs.

**Task.** Cut manual sales-ops work without boiling the ocean.

**Actions.**
- Rebuilt the **CRM UI** around daily workflows (pipeline clarity > feature tourism).
- Prototyped **AI document analysis** for the docs that blocked deals.
- Designed **automation** for repetitive handoffs instead of asking people to “try harder.”

**Result.** About **~40% less** manual sales-ops work. Hook: *I optimize the path the team walks every day, not the demo reel.*

**If they dig:** what you measured; what you refused to automate; how you handled wrong AI extractions.

---

## 4. Failed product decision + lesson

**Pick a true one.** Candidate (edit with your real details if needed):

**Situation.** SonicStream — sonic-enhanced LLM streaming. Felt clever on the subway: add another sensory channel to streamed tokens.

**Task.** Ship a memorable AI UX experiment.

**Actions.** Built the streaming + audio mapping UI, then tried it as a chat differentiator.

**Result.** Mostly **useless for AI chat**. Rescued the idea by pivoting to public-domain books as “listen to the sine waves of the text,” which is neat — and still not a company.

**Lesson.** Novelty ≠ job-to-be-done. I now force a user outcome sentence before an interaction gimmick. FlyingMinds/IvyOdds only keep features that change mastery, odds clarity, or time saved.

**Alternate failure (if you prefer):** shipping curriculum breadth before grading reliability — would have scaled distrust. Eval harness first was the correction.

---

## Delivery tips

- Open with the **metric** in the first 15 seconds.
- Name **one decision** you would reverse with today’s information.
- End with what you’d do in **their** product, in one sentence.