Skip to content

README.md

Latest commit

 

History

History
129 lines (88 loc) · 3.09 KB

README.md

File metadata and controls

129 lines (88 loc) · 3.09 KB
Raw

SignalBoard

A lightweight feedback triage tool that turns messy seller feedback into structured insights and flags high-risk issues for review.

Built as a minimal proof-of-concept demonstrating structured AI-assisted decision support tool for commerce platforms and feedback triage with human-in-the-loop safeguards.

Why This Exists

Sellers receive feedback everywhere:

  • Reviews
  • Support tickets
  • Refund notes

The problem isn’t reading feedback.
It’s knowing what matters.

SignalBoard extracts structured friction signals from unstructured text, scores confidence, and routes uncertain or high-risk cases to human review.

It separates noise from operational risk.

What It Does

  1. Accepts pasted feedback (reviews, tickets, refund notes)
  2. Uses DeepSeek (OpenAI-compatible API) to extract structured signals
  3. Applies deterministic review rules locally
  4. Aggregates results into a visual summary:
    • Top issue
    • High churn risk percentage
    • Payment-related flags
    • Human review required percentage
  5. Displays detailed per-item insights with confidence scoring

System Design

AI Layer

  • Extracts structured JSON signals from messy text
  • Returns:
    • friction_type
    • sentiment_score
    • churn_risk
    • payment_issue
    • chargeback_risk
    • confidence_score

Deterministic Rules Layer

After model output, the system:

  • Flags low-confidence outputs
  • Escalates payment-related negative sentiment
  • Routes medium/high chargeback risk
  • Handles malformed responses safely

The model assists.
The system decides.

Trust & Safety Considerations

  • LLM outputs are probabilistic.
  • Confidence scoring is required.
  • Payment-related issues are treated conservatively.
  • Low-confidence predictions trigger human review.
  • The UI never crashes on malformed model output.

This design assumes AI is an assistant, not an authority.

Tech Stack

  • Next.js (App Router)
  • TypeScript
  • Tailwind CSS (light/dark mode)
  • DeepSeek API (OpenAI-compatible)
  • Vercel (deployment)

Deployment

Deployed on Vercel using a server-side API route to protect API keys.

Environment variables are configured in the Vercel dashboard.

Product Tradeoffs

  • Batch processing vs real-time streaming
  • False positives vs false negatives in payment flagging
  • Strict JSON enforcement vs model flexibility
  • Automation vs human oversight

Future Directions (Not Included in MVP)

The current version is intentionally constrained to a single-page proof-of-concept demonstrating core signal extraction and triage logic.

  • Trend detection over time
  • Seller dashboard mode
  • Integration with ticketing systems
  • Risk prioritization scoring
  • Model comparison or ensemble scoring

Why This Matters

Commerce platforms operate at scale.

Surfacing the right signals early helps:

  • Reduce churn
  • Prevent chargebacks
  • Improve checkout experience
  • Protect seller trust

This project demonstrates AI-native product thinking:

  • Structured outputs
  • Guardrails around model uncertainty
  • Human-in-the-loop design
  • Clear separation between AI extraction and business logic