FORGE — Business Intelligence Engine

Raw input. Structured output. No noise.

**Live demo → https://forge-swarnimtiwari-projects.vercel.app/

FORGE is not a chatbot. It is a deterministic automation pipeline: unstructured business text goes in, validated structured data comes out. Every result is rendered in a purpose built UI — tables, risk panels, decision trees, action item trackers. Not a chat bubble.

Built as part of AI & Management research into applied LLM automation for enterprise knowledge work.

---

What It Does

Eight workflows, each processing a different class of business document:

Workflow	Input	Output
📄 Document Intelligence	Any business document	Type, summary, entities, action items, confidence score
🧾 Invoice Processor	Invoice text	Structured line items, totals, validation flags
👤 Customer Router	Support message	Category, routing, SLA, drafted auto-response
✉️ Email Composer	Bullet points or context	Complete professional email, CC suggestions
📋 Meeting Notes	Transcript or rough notes	Decisions, action items with owners, blockers
⚖️ Contract Reviewer	Contract text	Risk level, red flags, missing clauses, negotiation actions
🔬 Research Intelligence	Paper abstract or report	Methodology, claims, limitations, real-world applications
📊 Business Case Analyzer	Proposal or initiative	Recommendation, financial analysis, risk matrix

All outputs follow strict JSON schemas. The model is forced to return structured data — no markdown, no explanation, no deviation.

---

Works Instantly — No API Key Required

FORGE ships with a Simulation Mode built in. Every workflow returns realistic, detailed pre-built responses with authentic latency. Any recruiter or reviewer can open the deployed URL and see the full product working in three seconds — no sign-up, no key, no friction.

To use real AI (free options available):

Provider	Cost	Setup
Simulation	Free (built-in)	Nothing — works now
Groq (Llama 3.3 70B)	Free tier	console.groq.com — 2 min, no credit card
Google Gemini	Free tier	aistudio.google.com — Google account
Anthropic Claude	Paid	console.anthropic.com

---

Architecture Decisions Worth Reading

1. Provider Inversion

The entire workflow layer is decoupled from the AI provider. workflowDefs.jsx defines what to do. api.js decides how to call the model. Switching from Groq to Gemini to Anthropic is one dropdown change in Settings with zero workflow code changes. This is dependency inversion applied at the API boundary.

Workflow Layer  →  callAI(provider, workflowId, input)
                          ↓
                 api.js dispatches to:
                 Simulation | Groq | Gemini | Anthropic

Each provider has a different API contract:

• Groq: OpenAI-compatible /chat/completions with Authorization: Bearer
• Gemini: Google's contents[] format, key in URL
• Anthropic: messages[] format with x-api-key header

All three handled behind one function signature. Adding a fifth provider means adding one case in a switch statement.

2. React Inner-Component Anti-Pattern (Solved)

Tab components (Dashboard, WorkflowsTab, etc.) are called as {Dashboard()} — plain function calls and not rendered as <Dashboard />.

If you define a component inside App and render it as JSX, React creates a new component type on every App re-render (new function reference = new type). React unmounts the old instance and mounts a fresh one. All DOM state including focused inputs is destroyed. Every keystroke triggered this cycle, making every input unusable.

The fix is documented in code comments so it's a genuine learning artifact, not a hidden patch.

3. Role-Based Access Control (RBAC)

Four roles with enforced permission boundaries:

Role	Workflows	Pipeline	Audit Log	Analysis	Users	Settings
Admin	All 8	✓	✓	✓	✓	✓
Manager	All 8	✓	✓	✓	✓	—
Analyst	Doc/Inv/Contract/Research/BizCase	—	✓	✓	—	—
Viewer	—	—	—	—	—	—

The UI gates itself dynamically: tabs disappear, workflow cards lock, the nav rebuilds. Switch between users from the header dropdown available to every role including Viewer. No page reload required.

4. Pipeline Chaining

Any two workflows compose automatically. Workflow A's JSON output is serialized to structured text and fed as Workflow B's input. Suggested chains:

• Meeting Notes → Email Composer — extracts decisions, drafts the follow-up
• Research Intelligence → Business Case Analyzer — converts a paper into a proposal
• Contract Reviewer → Email Composer — identifies red flags, drafts the negotiation email

5. Log Intelligence (Analysis Tab)

Natural language queries over the audit log. Ask anything — "show me all high-risk contracts", "who ran the most jobs today?", "are there any anomalies?" — the AI reads the last 60 log entries and returns a structured analysis with relevant entries, patterns, and actionable insights.

---

File Upload

Every workflow accepts file uploads (.txt, .md, .csv, .json, .html). Drop a contract PDF converted to text, a meeting transcript export, or a research paper abstract — the content loads directly into the input. No parsing library needed since all supported formats are plain text.

---

Project Structure

forge/
├── index.html              # Single HTML shell
├── vite.config.js
├── package.json            # 3 runtime deps: react, react-dom, recharts
├── .env.example
└── src/
    ├── main.jsx            # React DOM mount — 6 lines
    ├── theme.js            # All design tokens + CSS keyframes
    ├── api.js              # Multi-provider AI abstraction
    ├── workflowDefs.jsx    # 8 workflow configs + output renderers
    ├── App.jsx             # Root — all state + 7 tab views
    └── components/
        └── Atoms.jsx       # Badge, Dot, RunBtn, ToggleSwitch, etc.

Adding a new workflow: one object in WORKFLOWS array in workflowDefs.jsx. Dashboard metrics, pipeline selectors, settings toggles, RBAC lists, and log prefixes all update automatically.

---

Quick Start

# Clone
git clone https://github.com/YOUR_USERNAME/forge.git
cd forge

# Install (3 runtime dependencies)
npm install

# Run — works immediately in Simulation Mode
npm run dev
# → http://localhost:5173

Add a real AI provider (optional):

cp .env.example .env
# Add VITE_GROQ_API_KEY — free at console.groq.com

Or skip .env — open the app, go to Settings → API Provider, paste your key, click Save.

---

Deployment

Vercel (recommended — free)

npm i -g vercel
vercel

Add VITE_GROQ_API_KEY in Vercel dashboard → Project Settings → Environment Variables.

Netlify

Connect GitHub repo. Build command: npm run build. Publish directory: dist.

---

Tech Stack

	Technology	Why
Framework	React 18 + Vite	Fast dev server, clean component model
Charts	Recharts	Composable, React-native data viz
AI	Groq / Gemini / Anthropic / Simulation	Provider-agnostic by design
Styling	Inline CSS-in-JS + design tokens	Zero class collisions, portable
Fonts	Barlow Condensed + JetBrains Mono	Loaded via Google Fonts

3 runtime dependencies total. No CSS framework. No state management library. No component library.

---

Domains This Touches

Applied AI · LLM Structured Output Enforcement · Multi-Provider API Design · Role-Based Access Control · Business Process Automation · Frontend Architecture · Design Systems · Enterprise Software Patterns · Prompt Engineering · Synthetic Data for Demo Environments

---

License

MIT — use it, fork it, extend it.