Demos and Experiments —
Clearly Labeled
What you see here is the method itself: synthetic data walking the full chain of problem → workflow → output → limits.
Product Documentation → Searchable Knowledge Base
◈ Synthetic demo data
Problem & hypothesis
Product parameters live scattered across PDF catalogs, Excel spec sheets, and certification scans. Sales keeps hunting through folders to answer customer questions. Hypothesis: once structured with source markers, retrieval drops from minutes to seconds — and every answer traces back to its source document.
Workflow
product files → field extraction → source markers → human review → searchable KB
Interactive demo (click the tabs to switch samples)
Three synthetic Q&A pairs. Watch the second one: when information is missing from the docs, the output says “unknown” and flags it for addition — the model never guesses.
Q: Heating range and certifications for this device?
A: Heating 38–55°C (source: spec sheet v3, p.2)
Certifications: CE, RoHS (source: cert scan 2024-03)
Q: Which languages does the on-device UI support?
A: Unknown — not found in current docs; the model does not guess. Flagged for addition to the spec sheet.
Q: What are the packaging dimensions and unit weight?
A: 32 × 22 × 11 cm, 2.4 kg (source: packaging spec 2024-06)
All inputs and outputs are synthetic data; this demo explains the method, not a client production deployment.
Known limits
- OCR accuracy depends on file quality; poor scans raise extraction failures
- Fields do not enter the KB without human review
- Information absent from source docs is marked "unknown" — never guessed by the model
Turning this into your pilot
Bring your real (redacted) product docs → discovery assesses file quality → a minimal pilot is scoped.
Inquiry Email → Classification & Reply Draft
◈ Synthetic demo data
Problem & hypothesis
Export inquiries arrive in mixed formats and languages; manual triage and drafting eat sales hours. Hypothesis: auto-classification + structured extraction + reply drafts (human-approved before sending) can cut per-inquiry handling time by half without losing human oversight.
Workflow
inbox → field extraction (product/qty/market) → classification → reply draft → human approval → send
Interactive demo (click the tabs to switch samples)
One standard inquiry, one complex inquiry. Watch the second one: anything beyond the AI draft boundary routes straight to a human — no auto-reply.
In: Hi, we need 500 units of your massage device for our EU market. What certifications do you have? Lead time?
Class: Inquiry · High priority (explicit qty+market)
Extracted: product=massage device | qty=500 | market=EU | focus=certs+lead time
Draft: Dear…, thank you for your inquiry. Our massage devices are CE/RoHS certified… (AI draft, pending human confirmation)
In: 3 units from the last batch failed heating — we demand compensation and a technical explanation. (synthetic complaint sample)
Class: Complaint · technical dispute
Action: No auto-draft — routed to a human
Reason: compensation terms and technical commitments exceed the AI draft boundary; a person must decide
All inputs and outputs are synthetic data; human approval gates every outbound action.
Known limits
- Nothing is sent automatically — approval gates every outbound email
- Pricing and discounts are never decided by AI; it only assembles the human-provided terms
- Complex inquiries (customization, complaints, technical disputes) route to humans
Turning this into your pilot
Bring redacted historical inquiries → discovery aligns classification rules → a small trial runs against your manual baseline.
Want to see your real data in the loop?
Bring one redacted sample (or a synthetic stand-in) and let's talk about a pilot.