Twelve working demos, one architecture pattern
Each tab is a small, self-contained proof of concept built around a recurring enterprise problem: unstructured documents (or plain-language requests) in, structured or governed output out. The pattern underneath is the same one used for Bedrock delivery — ingest, extract, reason, structure, audit.
| Demo | What it shows |
|---|---|
| PII Masker | Detects and redacts sensitive data in a document, with a classification log |
| Invoice Extractor | Turns a photographed or scanned invoice into structured, exportable data |
| Document Q&A Agent | Answers questions over an uploaded document, citing where the answer came from |
| Contract Comparer | Flags material differences between two versions of a document |
| Cloud & Compliance Advisor | Reviews an architecture description against NIST CSF / SOC 2 controls |
| Call & CRM Sync | Transcribes a live call, extracts key fields, and drafts the CRM record update |
| Call Recorder | Records or pastes a call transcript and saves it to S3 with a download link |
| Guardrails Demo | Same prompt, with and without a Bedrock Guardrail — shows exactly what gets blocked and why |
| Document Router | Classifies an uploaded document, extracts key fields, and routes it to the right team |
| Model Comparison | Same prompt across Haiku / Sonnet / Opus — latency, tokens, and cost side by side |
| Dashboard Builder | Throws several unrelated files at Claude and builds a KPI/chart dashboard from what it finds |
| As-Built Generator | Turns an architecture diagram into a structured as-built document, with open questions flagged instead of guessed |
Use the tabs on the left to try each one.
PII / PHI Masker
Paste text or upload a document. The model identifies sensitive fields, masks them, and returns a classification log — the pattern behind NYDFS / NIST-aligned data handling. A file upload runs on the real AWS pipeline (S3 → Textract → Bedrock → S3/DynamoDB); pasted text runs a direct Claude call for a quick preview.
Invoice / Receipt Extractor
Upload a photo or scan of an invoice or receipt. The model reads it and returns structured line items you can export as a spreadsheet.
Document Q&A Agent
Upload a document once, then ask it questions. The agent grounds every answer in the document and tells you when something isn't covered — the reasoning core of a RAG pipeline, normally backed by OpenSearch or Pinecone.
Contract / Policy Comparer
Upload two versions of the same document. The agent flags what materially changed — additions, deletions, and altered obligations — rather than a line-by-line text diff.
Cloud & Compliance Advisor
Describe an architecture — or paste a resource list — and get a review against common control frameworks, with concrete remediation steps.
Real-Time Call Transcription & CRM Sync
Speak into your mic (Chrome/Edge) or paste a transcript, then let the agent pull out the fields a rep would otherwise type manually — contact, pain points, next steps, sentiment, deal stage — and draft the CRM record update.
Call Recorder
Record a call live (Chrome/Edge mic) or paste a transcript, then save it straight to S3 with a short-lived download link — the "keep a copy on file" half of call handling, separate from the CRM field-extraction tab. A real deployment would swap the browser mic for streaming Amazon Transcribe and set a lifecycle policy on the bucket instead of a 5-minute link.
Guardrails & Responsible AI
The same prompt, run twice through the same deployed API — once with no guardrail, once with an Amazon Bedrock Guardrail attached that denies financial/medical/legal advice, blocks obvious prompt-injection attempts, and blocks or masks PII in the input. This is the governance conversation almost every regulated-industry buyer eventually asks for.
Without guardrail
With guardrail attached
Intelligent Document Router
Upload any document. The real AWS pipeline (S3 → Textract → Bedrock → DynamoDB) classifies it, extracts the headline fields, assigns it to the team that would actually handle it, and flags anything the model isn't confident about for human review — classify, extract, and route in one pass.
—
Extracted key fields
Model Comparison
The same prompt sent to Haiku, Sonnet, and Opus in parallel through the same deployed endpoint — latency, token usage, and estimated cost side by side. The single most common question in any Bedrock conversation is "which model do I actually need"; this makes the tradeoff concrete instead of theoretical.
Dashboard Builder
Throw a handful of unrelated files at it — invoices, reports, exported data, scanned forms, whatever's on hand — and the model reads all of them together, finds what's quantifiable, and builds a dashboard from scratch: no template, no predefined schema. Good for the "can it make sense of our messy data" question.
As-Built Generator
Upload an architecture diagram — a draw.io export, a whiteboard photo, a Visio screenshot, whatever you've got — and get back a structured as-built document: components, data flow, security notes, and (important) the things the diagram doesn't make clear, flagged as open questions rather than guessed at. Add project context below for a sharper read.
Keep Awake
Uses the browser's real Wake Lock API to stop this computer's screen from sleeping while this tab stays open. Genuine capability, one honest limitation: it only holds while the tab is open and in the foreground — switching tabs or minimizing releases the lock (a browser-enforced rule, not something this page can override), and it does not, by itself, stop Teams (or anything else) from showing you as "Away" if you step away from the keyboard — that's a separate, OS-level idle check unrelated to screen sleep.