DEMO_SCRIPT.md

Demo Script — llmtrace (3–4 min)

Recording checklist before you start:

• llmtrace seed has been run (fresh DB)
• GEMINI_API_KEY set, server running on port 8080
• Browser open at http://localhost:8080 (or live Vultr URL)
• Terminal visible alongside browser (split screen)
• Font size bumped (browser 125%, terminal 16pt)

---

Section 1 — The Problem (30 sec, talking head or voiceover)

"Every team building with LLMs eventually gets surprised by a bill. Helicone and Langfuse tell you that spend went up. But nobody tells you which deploy caused it. That's what llmtrace does."

---

Section 2 — The Dashboard (45 sec)

On screen: browser at the dashboard URL.

"This is the llmtrace dashboard. Thirty days of LLM call data across three API keys. prod-frontend in red, two control keys in grey."

Point to the chart.

"You can see everything was flat — about four dollars a day — until May third. Then prod-frontend spikes to twenty dollars overnight and stays there. The amber line is a deploy marker. PR 129 landed at 14:05 UTC. The spend broke loose eleven minutes later."

Scroll down to the anomaly cards.

"The detector flagged it at 28 sigma above the seven-day baseline. Eight dollars and twenty-four cents above expected on the first day alone."

---

Section 3 — The Agent (90 sec, the hero moment)

On screen: anomaly card for 2026-05-03.

Click "Investigate →".

"I'm clicking Investigate. This kicks off an autonomous agent — powered by Gemini — that queries the call ledger with three tools."

Wait for tool lines to stream in. Narrate as they appear:

"First tool: model distribution around the anomaly date. It sees two models in use on prod-frontend — haiku and sonnet."

"Second tool: deploy lookup in a four-hour window around May 3rd. One deploy found: PR 129, 'switch summary endpoint to claude-sonnet'."

"Third tool: prompt model diff. Same prompt fingerprint — same code path — but before the deploy it was ninety-one percent haiku. After: eighty-nine percent sonnet. And call volume jumped fifty-eight percent — the new prompt has a retry loop."

Wait for the attribution to stream in.

"There's the attribution. Zero-point-nine-five confidence. PR 129 caused it. The model swap plus the retry loop multiplied cost by four-point-two times."

---

Section 4 — CLI (30 sec)

Switch to terminal.

GEMINI_API_KEY=xxx llmtrace analyze --days 30

"Same investigation from the terminal — no browser needed. Pipes into CI, scripts, Slack bots. Wherever you run post-deploy checks."

Let it stream the first few lines, then cut.

---

Section 5 — Close (20 sec)

"llmtrace is self-hosted, open-source, and deploys in two minutes on a six-dollar Vultr VM. No SaaS dependency, no data leaves your infra.

Every team at this hackathon is burning tokens right now. When your bill surprises you next week — llmtrace tells you which deploy to blame."

Show: GitHub repo URL on screen.

---

Tips

• Keep Section 3 as the longest — the live streaming is the most visual.
• If the agent hits a rate limit and shows the retry message, that's fine — it demonstrates real behaviour, don't cut it.
• Record at 1080p, export at 1080p30. lablab.ai accepts YouTube/Loom links.
• One take is fine; the streaming output is different each run (model output varies).