01
Agentic AI Full-Stack MCP

AI Agent Operations Platform

A modular, production-grade system for deploying, orchestrating, and managing LLM-powered agents across business workflows — with a full client portal and admin dashboard.

The Problem

Most AI agent implementations are brittle: a single agent, a single workflow, no visibility into what's happening, and no structure for when it breaks. Businesses deploying AI at any real scale need something more — a system that orchestrates multiple agents, enforces output quality, keeps humans in the loop at the right moments, and gives non-technical stakeholders a clear view of what the agents are doing.

What Was Built

A full-stack AI operations platform with six specialized agents — research, copywriting, advertising strategy, offer creation, landing page generation, and sales management — each with defined input/output schemas, handoff protocols, and dependency resolution. A client portal exposes agent outputs and pipeline status to end users. An admin dashboard provides cross-client analytics, pipeline monitoring, and quality scoring.

The orchestration layer handles the logic that makes multi-agent systems actually work: checking whether upstream dependencies are complete before triggering downstream agents, resolving conflicts between agent outputs, and routing tasks to the right agent based on context.

Key Design Decisions
Human-in-the-loop checkpoint architecture
High-stakes outputs (client-facing copy, ad creative, sales enablement) route to a human review step before delivery. The system uses structured output scoring to determine confidence — low-confidence outputs are flagged automatically. This isn't a UX afterthought; it's designed into the pipeline logic from the start, because users who can't predict when an agent will ask for review don't trust the system.
MCP (Model Context Protocol) for modular tooling
Instead of building tool integrations directly into agent prompts, the platform uses MCP server architecture to give agents access to external data sources and APIs as composable modules. This keeps agent context clean, makes individual tools replaceable without touching agent logic, and enables capability expansion without re-engineering the core system.
Quality validation as a first-class system
A 58-point validation engine runs on every agent output before it surfaces to users. Structured scoring across output type, format compliance, and content quality — not just a vibe check. Outputs below threshold are recycled back into the agent with specific failure context rather than surfaced as-is. This is what makes the system usable in client-facing contexts.
Client data isolation at the schema level
Every table, query, and agent output is scoped to a client_id via Supabase Row-Level Security. Multi-tenant isolation is enforced at the database layer, not the application layer — so there's no path for one client's data to surface in another client's context regardless of application logic.
Stack
  • Next.js — frontend + API routes
  • Node.js — server logic, pipeline orchestration
  • Supabase — Postgres, Auth, Row-Level Security
  • Vercel — deployment, cron triggers
  • Claude API — all LLM inference
  • Python — automation scripts, data pipelines
  • MCP — modular agent tooling
Capabilities
  • 6+ specialized agents with full orchestration
  • Multi-tenant client portal + admin dashboard
  • BYOK encrypted API key management
  • Webhook integrations with third-party platforms
  • 58-point quality validation engine
  • Automated pipeline triggers via Vercel cron
  • Centralized telemetry and activity logging
Request a walkthrough →
02
OSINT Geospatial Data Viz

WorldView

Real-time geospatial intelligence dashboard — live flight tracking, satellite orbital paths, GPS jamming zones, and seismic events, layered on a 3D globe. Seven free public APIs. No credentials required.

The Problem

Open-source intelligence (OSINT) data is scattered across dozens of public APIs, each with its own format, rate limits, and interface. Analysts and researchers who want a unified view of geospatial signals — what's flying where, what's jamming GPS, where tectonic activity is spiking — have no single tool that brings it together without requiring clearances, subscriptions, or custom infrastructure.

What Was Built

A browser-based 3D geospatial dashboard built with CesiumJS that aggregates seven live public data streams: real-time commercial and military aviation (ADS-B), satellite orbital tracking (TLE data), GPS jamming and spoofing alerts, seismic event feeds, and additional geospatial intelligence layers. Everything renders on a photorealistic 3D globe with full camera control and layer toggling.

Built entirely with Claude Code — the architecture, data pipeline design, and rendering logic were developed through an AI-assisted workflow that's documented in full on the blog and YouTube channel.

Key Design Decisions
No-auth public API architecture
Deliberately constrained to free, no-auth APIs — not because of cost, but because it proves the capability ceiling without infrastructure dependencies. The tool works in any browser, for anyone, immediately. That constraint forced creative data sourcing and efficient client-side processing.
3D globe over flat map
Geospatial signals like satellite orbital paths and GPS jamming zones are distorted by flat projections. CesiumJS's 3D rendering preserves the actual geometry of the data — orbital ellipses look correct, jamming radius footprints scale accurately. It's not an aesthetic choice; it's the technically correct representation for this data type.
Stack
  • CesiumJS — 3D globe rendering
  • JavaScript — data pipeline + UI logic
  • 7 public APIs — ADS-B, TLE, GPS jamming, seismic + more
  • Claude Code — AI-assisted development
Data Layers
  • Live commercial + military aviation (ADS-B)
  • Satellite orbital tracking (TLE)
  • GPS jamming and spoofing zones
  • Real-time seismic events
  • Additional geospatial intelligence layers
Read the full breakdown → Watch the build →
03
Community SaaS Content

The AI Stack

A community and resource platform for people building seriously with AI — agents, pipelines, workflows, tools. Practical systems from someone actively deploying them.

The Problem

Most AI content teaches concepts. Almost nothing teaches deployment — the actual decisions made when a multi-agent pipeline breaks at 2am, or when a client's data starts bleeding into the wrong agent context, or when you're choosing between three architectures that all theoretically work. The people who know this stuff aren't writing about it.

What Was Built

A subscription community hosted on Whop, paired with a YouTube channel (@KareemxAI), that documents real AI builds in real time. Every project — the agent platform, WorldView, workflow tools — gets broken down into architecture decisions, what broke, what shipped, and why specific choices were made. Members get first access to new builds and the technical walkthrough before anything goes public.

The content side grew out of earlier work applying behavioral psychology to social media content design — scaling platforms to over 50 million combined views by understanding how attention, reward, and engagement actually work on algorithmic feeds. That same behavioral lens now informs how the AI content is structured and delivered.

Key Design Decisions
Build-in-public as the product
The most valuable content isn't polished tutorials — it's the unfiltered record of how a real system gets built. Architecture debates, wrong turns, debugging sessions, deployment failures. Members aren't watching someone who figured it out; they're watching it get figured out in real time. That's the format that builds genuine technical understanding.
Behavioral psychology applied to content design
50M+ views isn't a distribution fluke. It comes from applying cognitive science principles — variable reward schedules, curiosity gaps, cognitive load management — to how content is structured and paced. The psychology background isn't separate from the AI work; it's what makes the way the AI work gets communicated actually land.
Platform
  • Whop — community + subscription infrastructure
  • YouTube — @KareemxAI, build documentation
  • 50M+ combined views across content platforms
What members get
  • Real agent workflows and pipeline breakdowns
  • Claude Code techniques for serious projects
  • Architecture decisions from production builds
  • First access to new projects before they go public
Join The AI Stack → Watch on YouTube →
Currently Available

Open to the
right role.

Looking for roles at the intersection of agentic AI and human-centered design — where the work is building systems that real people actually trust and use.

UX for AI / Agentic AI AI Product AI Researcher Solutions Engineer Founding Team
Book a Call View Resume LinkedIn