docs/DEMO_SCRIPT.md

Demo video script — HCC-MEAT Coder (60–90 seconds)

Target length: 75 seconds. Shoot at 1920×1080 with the browser zoomed to 110% so the text is readable on a recruiter's tiny laptop screen.

Tools: OBS Studio (free), or QuickTime screen-record on macOS. Narration recorded separately and mixed — no live audio.

---

Pre-roll (0:00–0:03)

Visual: title card, dark slate background.

HCC-MEAT Coder CMS-HCC v28 · MEAT-justified · calibrated · RADV-framed Research demo · synthetic data only

---

0:03–0:15 — Why this exists

Visual: cut to the three-panel UI with no note loaded yet.

Narration:

Medicare Advantage plans pay risk-adjustment coders to capture every HCC the chart supports — but if the chart doesn't back it, CMS takes it back in a RADV audit. So coders read notes and look for MEAT: Monitor, Evaluate, Assess, Treat. The standard.

---

0:15–0:30 — Load a case, run the pipeline

Visual:

1. Click the gallery dropdown, pick "Straightforward DM2 with chronic complications".
2. Results populate. Camera lingers on the suspect list.

Narration:

Paste a note — or load an example — and the pipeline returns a ranked suspect list. Each HCC shows its RAF weight, a calibrated confidence, and a MEAT coverage pill. HCC-18, diabetes with chronic complications: all four MEAT categories verified, LOW over-coding risk.

---

0:30–0:45 — Drill in, show the MEAT quotes

Visual:

1. Click HCC-18 in the center panel.
2. Right panel opens. Scroll gently through the four MEAT categories.
3. On the left, the note shows the four quotes highlighted by category color.

Narration:

Every piece of MEAT evidence is a byte-exact quote from the note — not a summary, not a paraphrase. The verifier drops any quote it can't find in the source. Over-coding risk is a five-component transparent heuristic — named reasons, not a black box.

---

0:45–0:58 — Show a negative case

Visual:

1. Gallery dropdown → "Family history of malignancy".
2. Suspect list is empty.

Narration:

Family history alone doesn't code. Neither does "ruled out", "denies", or workup mentions. The pipeline filters those before linking, so the coder doesn't have to chase a suspect list full of false positives.

---

0:58–1:10 — Tabs for metrics + auditability

Visual:

1. Click the Metrics tab. Linger on the macro-F1 stat and the per-HCC table.
2. Click Subgroups. Linger on note-length slices.
3. Click Calibration — the placeholder panel is visible ("not fit yet").
4. Click About.

Narration:

Every number on screen is reproducible from the eval harness. Subgroup slices by note length and scenario tag are in the UI, not hidden in a notebook. Calibration is explicit — if the model isn't calibrated yet, the tab says so.

---

1:10–1:18 — Close

Visual: back to the coding tab; scroll the suspect list to hero the final frame.

Narration:

Built for a 2026 healthtech portfolio. Public data, single laptop, HF Docker Space. Repo and demo links in the description.

---

End card (1:18–1:22)

Visual: static text.

github.com/chaaarlieyap/hcc-meat-coder huggingface.co/spaces/charlesyapai/hcc-meat-coder

---

Notes for the recording session

• Close every Chrome tab except the demo. Disable any browser extension that could overlay UI (1Password, Grammarly, LastPass).
• Use Cmd+Shift+P in DevTools → "Show rulers" to confirm the 16:9 aspect before recording.
• Pre-populate the gallery: run scripts/generate_meat_templates.py so the gallery has at least the 10 staple cases.
• Do two silent passes first to lock timing before recording narration.
• Final export: H.264 MP4, 30fps, 1080p. Target file size < 20 MB so it embeds cleanly in the Space README.