> Initializing system...
Agentic AI & Automation Engineer

I build AI agents that run your business operations.

Who I Am // About

Not a demo.
A working system.

I'm a computer science student who builds agentic automation systems — multi-agent workflows, RAG pipelines, and natural-language interfaces — using n8n, MCP, and modern LLMs. Every project on this page is a live system you can watch run in the demo videos, not a mockup or a slide deck.

I'm looking to bring that same approach to real business problems: turning the email, calendar, and document work that eats up a team's day into something an agent handles end-to-end.

Node 01 // Intake

Trigger &
Intake.

Every system starts at a trigger — a Telegram message, a webhook call, or a schedule. Payloads are normalized into one shape before anything reaches an agent.

Node 02 // Orchestrate

Agent
Orchestration.

One orchestrator agent reads intent, then hands off over MCP to a specialized sub-agent — Email, Calendar, or Web Search — each running its own isolated memory.

Node 03 // Retrieve

Context
Retrieval.

For document-heavy tasks, content is chunked and embedded with NVIDIA NeMo, then stored in Pinecone. Each query retrieves the nearest matching vectors before generation.

Node 04 // Execute

Tool
Execution.

Once routed, sub-agents call real MCP tools — sending an email, creating an event, running a search — and every exchange is persisted to Postgres so context survives across sessions.

Projects // Live Builds

AI Systems
I've Built.

Independent Build 01 // RAG AI Agent

Document retrieval pipeline.

A fully automated ingestion-to-chat pipeline: files are pulled on a schedule, chunked, embedded, and upserted into a vector store, then served through a conversational agent that answers only from that retrieved context.

RAG AI Agent demo thumbnail
WATCH DEMO
3
Pipeline Stages
NeMo
Embedding Model
Pinecone
Vector Store
Business Impact
  • →Faster document retrieval — answers pulled from source docs in seconds, not manual search
  • →Always current — new files are indexed automatically, no manual re-work
System Architecture
  • Frameworkn8n
  • LLMOpenAI
  • Vector DBPinecone
  • EmbeddingsNVIDIA NeMo
  • Text SplitterRecursive Char.
Design Note

Ingestion and chat run as separate workflows. Re-indexing new documents is a one-click re-run of the ingestion flow, no redeploy needed.

Independent Build 02 // Email & Calendar Agent

Multi-agent Outlook automation.

A Telegram-triggered orchestrator delegates to two dedicated sub-agents over MCP: one manages Outlook email (triage, draft, reply, folder routing), the other manages calendar events, each with its own isolated conversation memory.

Outlook Email & Calendar Agent demo thumbnail
WATCH DEMO
32
Nodes Built
2
MCP Sub-Agents
GPT-5.1
Model
Business Impact
  • →Automated email management — triage, drafting, and replies without opening Outlook
  • →Natural-language calendar management — schedule, move, or cancel events from a chat message
  • →Reduced manual work — one Telegram thread replaces inbox and calendar app-switching
System Architecture
  • Frameworkn8n + MCP
  • LLMGPT-5.1
  • MemoryPostgres (per-agent)
  • IntegrationOutlook (OAuth)
  • TriggerTelegram
Design Note

Each sub-agent keeps its own Postgres memory table, keyed to the Telegram chat. Email and calendar context never bleed into each other.

Independent Build 03 // AI Executive Assistant

Multimodal Telegram assistant.

Accepts text, voice notes, or images from Telegram, routes all three through one orchestrator, then delegates to Email, Calendar, or Web Search sub-agents over MCP — replying back in text or generated voice.

AI Executive Assistant demo thumbnail
WATCH DEMO
3
Input Modes
3
MCP Sub-Agents
Voice
Output Option
Business Impact
  • →Unified AI assistant for employees — one interface for email, calendar, and research
  • →Works hands-free — voice in, voice out, no screen required to get an update
System Architecture
  • Frameworkn8n + MCP
  • LLMOpenAI
  • InputText / Voice / Image
  • OutputText + TTS Audio
  • ToolsEmail, Cal, Web Search
Design Note

Voice and image inputs are transcribed/analyzed and normalized into the same format as text before hitting the orchestrator, so routing logic only has to be written once.

Capabilities // Skills

What I Work With.

Orchestration
n8n MCP LangGraph Multi-Agent Systems Workflow Automation Event-Driven Architecture
AI & LLM
OpenAI Claude RAG Prompt Engineering Tool Calling Structured Outputs AI Evaluation Context Engineering
Data
PostgreSQL Pinecone ChromaDB Vector Search Embeddings Redis
Backend
Python FastAPI REST APIs Docker Linux Git OAuth
Cloud
Cloudflare Railway Render Docker Compose CI/CD Basics
Live Console

Execution Trace.

oussama@portfolio: ~/run_agent.py
$ n8n execute --workflow "email_calendar_agent"
> Trigger: Telegram message received
> [Node: Orchestrator] Classifying intent... → Email Agent
> [Node: Email Agent] Connecting via MCP (Outlook OAuth)
> [Node: Email Agent] Drafting reply...
> [Node: Postgres Memory] Session context saved
[Output] Tool call: send_email → success
Status: SENT. Awaiting next trigger.
$ _