KEY FEATURES
I built this to demonstrate agentic AI patterns—multi-step reasoning, tool use, and structured outputs—that are becoming essential in production AI systems. The goal was to create something functional, not just a toy demo.
​
-
Visible AI reasoning: Watch the agent's step-by-step thinking process as it classifies content, scans for sensitive information, evaluates quality, and generates improvements
-
Quality scoring: Content is evaluated on clarity, grammar, tone, and overall quality with scores from 1-10
-
Safety guardrails: Automatic detection of PII like Social Security numbers and credit card information
-
Smart content improvement: AI generates an enhanced version while preserving the original intent
-
Context-aware analysis: System detects content type (email, LinkedIn post, etc.) and applies appropriate standards
-
Real-time processing: All analysis is live via Claude API - nothing is hardcoded
HOW I BUILT THIS
Tech Stack
-
Frontend: Vanilla HTML, CSS, JavaScript
-
Backend: Python (Flask)
-
AI: Claude API (Anthropic)
-
Deployment: Render
-
Code Editor: Cursor (AI-assisted IDE) + Claude Code
My Process
I used a "Claude CTO" approach - creating a Claude project with custom instructions to act as an opinionated technical advisor. Before writing any code, I'd consult the CTO on architecture decisions, which helped me avoid rabbit holes and make smart tradeoffs. For implementation, I used Claude Code in Cursor to handle file creation, terminal commands, and debugging while I focused on product decisions.
Challenges Solved
-
UX timing issues: The loading animation needed to scroll into view immediately when analysis started, not after it completed. Required careful coordination between UI state changes and scroll behavior.
-
PII detection nuance: Initial implementation flagged all email addresses as PII, but an email in a signature isn't a security risk. Identified this as a V2 improvement: add a confirmation modal letting users choose what to redact.
-
Progressive disclosure: Results sections (Quality Scores, Guardrails, etc.) needed to stay hidden until analysis completed to prevent confusion from empty states.
-
Validation approach: Stress-tested with intentionally flawed content covering grammar errors, PII exposure, tone issues, and vague claims—all correctly identified and addressed by the system.
​
V2 Ideas
-
Strict PII mode toggle for compliance-heavy use cases (healthcare, finance)
-
Document type presets that adjust scoring weights (executive tone, marketing polish, casual)
-
User feedback on AI suggestions to improve recommendation quality over time
​