# LifeOS — 90-Second Demo Script

> One small model. One memory. Your whole week, handled — offline.

**Setup before recording:** Space freshly restarted (clean seed data), browser
at 1280px, `data/samples/flyer.pdf` and `data/samples/transactions.csv` on the
desktop. Optional but killer: have the OS wifi toggle visible in the menu bar.

---

### 0:00–0:10 — Cold open (the problem)
*Screen: LifeOS header with "● 100% LOCAL · Nemotron-3-Nano-4B · llama.cpp" chips.*

> "I'm a student. My grocery money, my training plan, and my subscriptions
> all leak in different apps — and I'm not sending my bank statements to a
> cloud AI. So I built LifeOS: a personal assistant that runs entirely on a
> 4-billion-parameter NVIDIA model, on a free CPU."

### 0:10–0:30 — 💰 Money: find the leaks
*Upload `transactions.csv` → subscription table appears → click "Review my subscriptions". Tokens stream live.*

> "I drop in three months of bank transactions. Deterministic code detects the
> recurring charges — then the model judges them against my income. It flags
> FitnessPal — untouched since March — and Crunchyroll. That's ~$32 a month
> back."

### 0:30–0:55 — 🍳 Kitchen: shop the deals, skip the repeats
*Upload `flyer.pdf` → deal chips appear → "Get recipe picks" → 3 recommendations stream in.*

> "Now dinner. I upload this week's grocery flyer — OCR and parsing happen
> locally. Watch the *why*: it picks chicken thighs because they're $2.49 a
> pound **and** because it remembers I've already had pasta twice this week.
> It knows I eat halal and chase protein. That's the shared memory working."

### 0:55–1:10 — 💪 Health: it knows my knee
*Log yesterday's leg day → "What should I do tomorrow?" → recommendation streams.*

> "I log my workouts here too. It suggests an easy morning run — because my
> 10K is in September, and its long-term memory knows my knee needs a rest
> day between runs. I never told it that today; it remembered."

### 1:10–1:25 — 🧠 The killer cross-domain ask
*Chat tab. Click the example chip: "Plan my week: under $80 groceries, high protein, run 3×". Answer streams, citing flyer prices, the cancel list, and the training plan.*

> "And this is the point of one assistant instead of three apps: one question
> that needs my food, my money, and my training — answered by a 4B model
> from one shared memory."

### 1:25–1:30 — The proof
*Toggle wifi OFF. Send one more chat message. It streams normally.*

> "Oh — and wifi's been off for this part. Off the grid. Built small."

*End card: LifeOS · Nemotron-3-Nano-4B · llama.cpp · 100% local · HF Build Small hackathon.*

---

**Beat notes:** keep each tab's streaming visible 3–4s before cutting; the
wifi-off beat must show the toggle AND the streaming response in one unbroken
shot — it's the credibility moment of the video.