MU Muhammad Usman
--:-- ISB
Portfolio / 2026 33.68° N, 73.04° E

Muhammad Usman.

Python Backend & AI Engineer

I work on the backend of AI products: getting messy data in, making retrieval accurate, and keeping everything running reliably at scale. Right now I'm at AAI Innovations (remote, Germany), owning backend features for CORA from technical design through production.

Scroll
0 Years shipping production AI & data systems
0k Regulations processed per month, autoscaled on Kubernetes
0 MCP servers behind one gateway
0 Countries shipped for: Germany, Portugal, Pakistan
(01)

Where I've worked

Click a role to expand

    • Build Python backend services, data pipelines and AI agent workflows for CORA, a regulatory intelligence platform, owning features from technical design through production deployment.
    • Engineered the automated document ingestion pipeline (document parsing, semantic chunking, LLM-based image captioning, vector embeddings) that processes 250k regulations per month, running on Kubernetes with KEDA-driven autoscaling.
    • Built an MCP gateway hosting 5 MCP servers over Streamable HTTP, letting customers' AI agents (Claude, Cursor) call individual product features, secured with OAuth 2.1 (CIMD) and per-user credit metering.
    PythonKubernetesKEDAMongoDBLangGraphVertex AIMCP
    • Built and maintained Apache Airflow DAGs and Python services for high-volume daily data ingestion, transformation and analytics, including vehicle homologation and safety-critical automotive data.
    • Improved throughput and reliability across these workflows, cutting manual fixes and keeping results correct and traceable under strict quality requirements.
    PythonApache AirflowETL
    • Built the company's first customer-facing AI chatbot on WhatsApp (Botpress plus custom code). Customers verify who they are, then get live account details from Airtable or file complaints with no staff involved. Thousands still use it.
    • Built Airtable automations and Node.js backend workflows, including a document parsing pipeline that streamlined document handling and made operations more efficient.
    Node.jsBotpressAirtableWhatsApp
    • Bachelor of Engineering in Mechatronics at the National University of Sciences & Technology, College of Electrical & Mechanical Engineering, Rawalpindi.
(02)

Selected work

The demos below are interactive

01 · AAI Innovations

CORA

An AI agent platform for regulatory compliance, covering 250k+ regulations from the EU, US, UN, ISO and others.

  • Engineered the ingestion pipeline: document parsing, semantic chunking, LLM-based image captioning and vector embeddings, processing 250k regulations a month on Kubernetes with KEDA autoscaling.
  • Agentic workflows with LangGraph and Vertex AI, traced and evaluated in LangSmith.
  • RAG on MongoDB Atlas Search with hybrid (vector + keyword) retrieval, for accurate, low-latency results across the full regulatory corpus.
ingestion.pipeline KEDA · Kubernetes
pods 1
throughput —
indexed 0

Drag the load. Pods scale up under pressure and back to zero when idle, the way KEDA runs it.

02 · AAI Innovations

MCP Gateway

The layer that lets customers plug specific product features into their own AI agents, without buying the full product.

  • One gateway hosts 5 MCP servers over the full Streamable HTTP protocol, working with clients like Claude and Cursor.
  • OAuth 2.1 with Client ID Metadata Documents (CIMD), which replaces Dynamic Client Registration, plus per-request authorization.
  • Tracks credit usage per user and routes each request to the existing product services, returning results to the agent.
mcp.gateway credits 8 

        

      


      
      

        

          
03 · Condoroo

          
WhatsApp support bot

          
The company's first customer-facing AI assistant. Customers verify who they are, then pull live account details or file complaints with no staff involved.

          
    
            
  • Botpress plus custom code, backed by Airtable as the live source of truth.
    • 
            
  • Node.js workflows and Airtable automations behind it, including a document parsing pipeline.
    • 
            
  • Thousands of customers still use it.
    • 
          

        


        

          

            
            

              Condoroo Support
              online
            

          

          

          
Illustrative conversation

        

      

    


    
    

      

        (03)
        
The toolbox

        
What I reach for

      


      

        

          AI & LLMs
          
LangGraphGoogle ADKMCP (Streamable HTTP)LangSmithRAGOpenAI & Anthropic APIsVertex AI

        

        

          Backend
          
FastAPIREST APIsOAuth 2.1

        

        

          Data & Storage
          
MongoDB Atlas SearchPostgreSQLApache AirflowETLWeb scraping

        

        

          Infrastructure
          
KubernetesKEDADockerGitHub ActionsCI/CDGit

        

        

          Languages
          
PythonJavaScript (Node.js)

        

      

    


    
    

      (04)
      

        Got messy data
        or an agent that
        needs a backend?
      


      

        

          
Write to me

          

          
        


        

          
Or leave a note

          
          
          
          
Opens your mail client with the message filled in.