ai / DEMO_SCRIPT.md
Aissugur Kurganbekov
Claude Opus 4.8 (1M context)
Add recruiter-facing docs: case study, architecture, interview notes, demo, CV, outreach
b2f4ef1
|
Raw History Blame Contribute Delete
1.92 kB

Demo Script β€” 60-second Loom

Setup before recording: backend running on :8000, frontend on :5173, browser on the Overview page, screen clean. Speak calmly; let the UI do the work.


0–10s β€” Problem

"Finance teams answer the same questions from invoices and contracts all day β€” what's the total, when's it due, what's the payment term. A normal chatbot will just make up a number. FinanceFlow only answers from the document, with citations β€” and refuses when it can't."

(On screen: the Overview hero + the four trust badges.)

10–20s β€” Upload / demo document

"I'll load the sample documents in one click."

(Click Try Demo. It seeds an invoice, a contract, and a payment note, and jumps to the Ask workspace.)

20–35s β€” Ask a supported question

"It already answered 'What is the total amount due?' β€” and here's the answer."

(The answer card shows the figure with a green "cited" badge. Point at it.)

35–45s β€” Show citations / evidence

"On the right is the exact source passage it used β€” document, chunk, and the highlighted line Total Amount Due $2,413.98. Every answer is verifiable."

(Hover/expand an item in the Evidence panel.)

45–55s β€” Ask an unsupported question β†’ abstention

"Now something the document doesn't contain β€” a social security number."

(Click the "Correctly abstains" chip. The amber card shows: "Not enough information in the uploaded document." with no citation.)

"It refuses instead of guessing. That's the whole point in finance."

55–60s β€” Evaluation + closing line (Click Evaluation β†’ Run.)

"And I measure that: a golden-dataset eval β€” 100% citation coverage, zero unsupported answers. Source-backed AI you can actually trust."


One-line close (if you have a spare second):

"Built with FastAPI, React, and a custom retriever β€” runs fully offline."