AgileX A.G.I.L.E. Portal

1 What This Is

The AgileX A.G.I.L.E. Portal is a working Cloudflare Workers application that demonstrates the full AgileX recruitment intelligence architecture. It is deployed at:

It proves that:

• Edge AI works — Llama 3.1 8B runs at the Cloudflare edge to parse CVs, score candidates, and draft communications
• Agents are autonomous — each A.G.I.L.E. agent has its own API endpoint and can process requests independently
• Microsoft 365 is the golden source — simulated Graph API endpoints show exactly how OneDrive, SharePoint, Teams, and Entra ID integrate
• Huly ATS is a metadata layer — the ATS stores minimum data (names, stages, links) while documents live in M365
• The whole stack costs less than R1,500/month — Cloudflare free tier covers Workers, D1, KV, and AI Gateway

2 Architecture

Data Ownership Principle

3 A.G.I.L.E. Agents

Five autonomous agents handle the recruitment lifecycle. Each runs as a function inside the Cloudflare Worker with access to D1, KV, and Workers AI.

How Agents Use Edge AI

4 Technology Stack

5 Project Structure

Total: 1,525 lines of TypeScript + HTML across 11 files.

D1 Database Schema (6 tables)

6 API Reference

Agent Endpoints

Data Endpoints

AI Endpoints

M365 Simulation & Huly Bridge

7 Microsoft 365 Simulation

Every M365 endpoint returns a "simulated": true flag and the exact Microsoft Graph API call it represents. In production, these swap to real Graph API calls with OAuth tokens from Entra ID.

8 Huly ATS Bridge

The portal demonstrates how AgileX connects to any Huly instance — a local server in a South African office, a Hetzner VPS in Europe, or scaled elastically as needed.

9 Edge AI — How It Works

Cloudflare Workers AI runs inference at the edge — in the data centre nearest to the request. No round-trip to a central API. Zero cold starts.

Models Used

Proven: CV Parsing at Edge

The Acquire agent was tested with a real CV. The Edge AI (Llama 3.1 8B) extracted:

{
  "name": "Dr. Nomsa Khumalo",
  "email": "nomsa.khumalo@gmail.com",
  "phone": "+27 82 555 1234",
  "city": "Johannesburg",
  "title": "Senior ML Engineer",
  "skills": "Python, TensorFlow, PyTorch, NLP, MLOps, Azure ML",
  "experience_years": 8,
  "current_company": "Discovery Health",
  "salary_expectation": 1300000
}

The agent then stored the candidate in D1, generated a simulated OneDrive link, and logged the activity — all in a single Worker request.

Getting Smarter

• Every CV processed can generate an embedding stored in Vectorize — the index grows, matching improves
• AI Gateway caching means repeated similar queries return instantly from cache
• Consultant feedback (overrides) logged in D1 → used to refine scoring prompts quarterly
• Cost decreases over time as cache hit rate increases (expect 30–50% after 3 months)

10 Setup & Deployment

Prerequisites

• Node.js 20+ and npm
• Wrangler CLI: npm install -g wrangler
• Cloudflare account (free tier is sufficient)
• Logged in: wrangler login

Local Development

# Clone and install
cd /Users/craig/Projects/AgileX/agilex-portal
npm install

# Initialise local D1 database
npm run db:init

# Start dev server (port 8788)
npm run dev

# Seed demo data
curl -X POST http://localhost:8788/api/seed

# Open portal
open http://localhost:8788

Note: Workers AI (edge inference) only works when deployed to Cloudflare. Locally, agents use fallback demo responses.

Deploy to Cloudflare

# Create remote D1 database (one-time)
wrangler d1 create agilex-portal
# Update wrangler.toml with the database_id

# Create KV namespace (one-time, or reuse existing)
wrangler kv namespace create CACHE
# Update wrangler.toml with the KV id

# Initialise remote D1 schema
wrangler d1 execute agilex-portal --remote --file=schema.sql

# Deploy worker + assets
npm run deploy

# Seed remote data
curl -X POST https://agilex-portal.<subdomain>.workers.dev/api/seed

Current Deployment

• Worker deployed: agilex-portal.ccodeimbila.workers.dev
• D1 database: agilex-portal (e60b6084-b097-4ce9-b07f-2952c319129a) — WEUR region
• KV namespace: CACHE (fa3239d69b08493da8e3acce91fca5e9)
• Workers AI: Llama 3.1 8B + BGE Base embeddings — tested and working
• Demo data seeded: 15 candidates, 8 vacancies, 10 clients, 8 activity logs

11 Using the Demo Portal

Dashboard Tab

Shows real-time metrics (candidates, placements, offers, agent actions), the agent activity feed, and pipeline stage counts. Data comes from the D1 database via the Leverage agent's dashboard endpoint.

A.G.I.L.E. Agents Tab

Click an agent card, then type commands in the chat input:

The right panel shows the raw JSON API response for every action — useful for developers integrating with the agents.

Pipeline Tab

Kanban board showing all candidates by stage. Cards show name, title, and AI score (colour-coded: green ≥80, amber ≥60, red <60).

M365 Layer Tab

Click any card to see the simulated Microsoft Graph API response. Every response includes the exact Graph API endpoint that would be called in production, plus realistic response shapes.

Huly Bridge Tab

Tests connectivity to the Huly ATS instance and shows the full integration map — endpoints, data model, and deployment options.

Edge AI Tab

Raw inference playground. Type any prompt and run it against Llama 3.1 8B at the Cloudflare edge. Also test embedding generation (768 dimensions via BGE Base) and view the available model list.

12 What's Next

Production Readiness (Weeks 2–4)

• Real Microsoft Graph integration — swap simulated endpoints for OAuth + real Graph API calls
• Huly WebSocket bridge — connect agents to Huly transactor for live ATS updates
• Cloudflare AI Gateway — add multi-model routing (Claude, GPT-4o, Gemini as fallbacks)
• Cloudflare Vectorize — persistent candidate embeddings for semantic search
• Resend integration — real email delivery from Engage agent
• Authentication — Entra ID SSO for portal access
• Custom domain — app.agilex.co.za via Cloudflare DNS

Intelligence Improvements

• Feedback loop — capture consultant overrides to refine scoring prompts
• RAG pipeline — Vectorize embeddings + D1 context for better candidate matching
• Scheduled agents — Cron triggers for daily candidate sourcing and weekly reports
• Multi-tenant — D1 row-level isolation per client workspace