Skip to content

Latest commit

 

History

4 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

AI Handoff Evaluator

"The most consequential question in AI product development isn't what AI can do but knowing exactly where it should stop."

A framework, position paper, and intelligent tool for human-AI task delegation. Built as a portfolio project demonstrating AI PM thinking: systems design, original research, and product execution in one.


What This Solves

As AI takes over more of the execution layer, teams face a recurring failure pattern: they either over rely on AI in high accountability contexts, or under use it out of vague fear. Both failures share the same root = the handoff between AI and human responsibility is implicit, undesigned, and ungoverned.

This project gives teams a structured method for drawing that line.


The Framework in One Sentence

Every task can be scored on two axes = Ambiguity (how judgment heavy is it?) and Accountability (how high are the stakes?). And the intersection of those scores determines whether AI, a human, or a designed handoff should own it.

The Four Zones

Zone Ambiguity Accountability Ownership Model
🤖 AI Zone Low Low AI executes, human spot-checks
🤝 AI-Assisted Zone Low High AI drafts, human certifies
🧭 Human-Led Zone High Low Human directs, AI assists
👤 Human Zone High High Human owns entirely

The 12 Dimensions

Tasks are scored across twelve dimensions that map to practical AI capability boundaries:

Ambiguity: Problem Framing · System Architecture · Product Judgment · Distributed Systems Thinking · Messy Requirements · Decision Under Uncertainty

Accountability: Security & Reliability · Infrastructure Strategy · Human Coordination · Domain Expertise · Business Constraints · Verifying Correctness


The Tool

A three step intelligent evaluator:

  1. Describe your task in plain language.
  2. Review Groq's LLaMA 3.3 70B analyzes it and pre scores all 12 dimensions, with a rationale and key handoff risk. You can adjust any score.
  3. Plan Get a zone classification, axis scores, and a step by step handoff protocol.

Project Structure

ai-handoff-evaluator/
│
├── src/
│   ├── main.jsx              ← React entry point
│   └── App.jsx               ← The full application 
│
├── docs/
│   ├── FRAMEWORK.md          ← The core framework document
│   ├── RESEARCH_PAPER.md     ← Position paper: "Where AI Stops"
│   └── SYSTEM_ARCHITECTURE.md ← Technical architecture notes
│
├── index.html                ← Vite HTML entry point
├── vite.config.js            ← Vite configuration
├── package.json              ← Dependencies and scripts
├── .env.example              ← 
├── .gitignore                ← Out of git
└── README.md                 ← 

Getting Started

1. Get a Free Groq API Key

Go to console.groq.com and create a free account. No credit card required. Copy your API key, starts with gsk_.

2. Clone the Repository

git clone https://github.com/YOUR_USERNAME/ai-handoff-evaluator.git
cd ai-handoff-evaluator

3. Install Dependencies

npm install

This installs React, React DOM, Vite, and the Vite React plugin. Nothing else. The project intentionally has zero runtime dependencies beyond React.

4. Configure Your API Key

cp .env.example .env

Open .env and replace gsk_your_key_here with your actual Groq key:

VITE_GROQ_API_KEY=gsk_your_actual_key_here

Note: You can also enter your key directly in the app's UI without setting up the .env file. It's never stored anywhere and is sent only to Groq's API over HTTPS.

5. Start the Dev Server

npm run dev

Open http://localhost:5173 in your browser. The app is live.


VSCode Setup

For the best development experience, open the project in VSCode and install these extensions:

  • ES7+ React/Redux/React-Native snippets — fast component scaffolding
  • Prettier — consistent code formatting
  • ESLint — catches common React errors before runtime
  • Vite — Vite-aware dev server integration

The project works out of the box with no VSCode configuration needed.


Build for Production

npm run build

Vite bundles the app into dist/. You can deploy the contents of dist/ to any static host like Vercel, Netlify, GitHub Pages, or Cloudflare Pages all work with zero configuration.

For Vercel specifically:

npm install -g vercel
vercel

Technology Decisions

The stack is intentionally minimal. The insight lives in the framework and the LLM prompt design, not in the infrastructure. Heavy dependencies would obscure that.

Layer Choice Why
UI React 18 Component model suits the multi step evaluation flow
LLM Groq LLaMA 3.3 70B Free, ~3s response time, OpenAI compatible API
Build Vite 5 Sub second HMR, zero config for React JSX
Styling Inline CSS + design tokens Zero runtime deps, fully portable, design system ready
State React useState No persistence needed; evaluation is session scoped

The Groq API call uses temperature 0.3 — low enough for stable, reproducible scoring but not so low that the model produces degenerate outputs. The system prompt includes explicit 1 and 5 anchors for each axis to prevent the model from clustering scores in the 2 to 4 range.


Docs

The docs/ folder contains the full intellectual layer of this project:

  • FRAMEWORK.md — The structured model: two axes, four quadrants, 12 dimensions, and how to use them.
  • RESEARCH_PAPER.md"Where AI Stops" is the position paper arguing this is the defining PM challenge of the next decade.
  • SYSTEM_ARCHITECTURE.md — Data flow, component structure, prompt design decisions, security model, and how this scales to production.

Portfolio Context

This project is designed to demonstrate three capabilities specifically relevant to AI PM roles:

Systems thinking = The framework shows the ability to decompose a complex sociotechnical problem (human-AI collaboration) into a structured, repeatable model.

Original argumentation = The research paper develops a non obvious, defensible point of view on a real problem in the field.

Product execution = The tool translates the thinking into something interactive and usable which's not just described, but built.


License

MIT. Use it, fork it, build on it.

Releases

Packages

Contributors

Languages