Skip to content
View umerjavaidkh's full-sized avatar
🏠
Working from home
🏠
Working from home

Block or report umerjavaidkh

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
umerjavaidkh/README.md

Umer Javaid

Senior AI/ML Engineer — Arabic-first LLM & Retrieval Systems at National Scale

10+ years building production systems, now focused on Arabic-language LLM and retrieval applications at national scale.

I build the thing Saudi Arabia's AI push needs most: Arabic-first, culturally-grounded conversational AI that actually works in production. I engineered the bilingual (Arabic + English) AI layer behind chat.dubai.ae, the official Dubai Government assistant, and built BilArabi — an Arabic educational RAG platform with a custom OCR-correction and 4-phase ingestion pipeline purpose-built for the challenges of Arabic text.


Core Work

  • Arabic NLP & RAG — bilingual retrieval, Arabic OCR correction, dialect-aware processing, government-grade terminology handling
  • GraphRAG & Knowledge Graphs — creator of Agentic GraphRAG (open source): knowledge graphs + multi-agent orchestration for reasoning over enterprise data
  • Production LLM Systems — RAG pipelines, multi-agent workflows, MCP routing, hybrid retrieval (BM25 + dense) with re-ranking, on Azure AI Foundry
  • National-Scale Delivery — UAE Pass digital identity (5M+ users), Digital Dubai, Emaar

Featured Project

agentic_graph_rag — Production-ready GraphRAG engine using Neo4j, LangGraph, and FastAPI for intelligent retrieval, Text-to-Cypher analytics, and secure graph-native RBAC.

Also shipped the Emaar / The Dubai Mall app, serving 1M+ users.

Portfolio · LinkedIn · GraphRAG (open source)


Future Vision

graph TD
    subgraph L6["L6: Experience & Client Interfaces"]
        UI_Slack["Slack / Teams"]
        UI_IDE["IDE / Claude Code"]
        UI_Web["Web Console / Partner Portal"]
        UI_WA["WhatsApp / Mobile Apps"]
    end

    subgraph L5["L5: Agent & Skill Plane (LangGraph + Claude SDK)"]
        AgentReg["Agent Registry"]
        SkillReg["Skill Registry"]
        Orchestrator["LangGraph Orchestrator"]
        ScopedMem["Scoped Memory (PG + Redis)"]
    end

    subgraph L4["L4: Governance & Control Plane (IntelliFlow Casbin PDP + AAT Minting)"]
        PEP1["PEP-1: Session Authenticator"]
        AAT["AAT Token Minter (RS256)"]
        Jurisdiction["Jurisdiction & Entity Resolver"]
        PDP["Policy Decision Point (Casbin / OPA)"]
        PEP2["PEP-2: Skill Approver"]
        RiskEngine["Risk Tier Engine (R0 - R5)"]
        ApprovalSvc["Temporal Approval Service"]
        CostLedger["Redis Cost & Budget Ledger"]
        AuditLog["Append-Only WORM Audit Log"]
    end

    subgraph L3["L3: Gateways & Enterprise Knowledge Brain"]
        subgraph ModelGW["Model Gateway (LiteLLM Router Core)"]
            PEP3["PEP-3: Model Enforcement"]
            PII["PII Redactor & Tokenizer"]
            Router["LiteLLM Class-Based Router"]
            SemCache["Redis Semantic Cache"]
        end

        subgraph ToolGW["Tool / MCP Gateway (Proxy)"]
            PEP4["PEP-4: Tool Enforcement"]
            Broker["Credential Broker (Vault)"]
            JurisInject["Server-Side Entity Injector"]
            BlastCap["Blast Radius & Rate Limiter"]
        end

        subgraph Brain["The Brain (Hybrid Storage Engine)"]
            PEP5["PEP-5: Data Filter Injection"]
            Neo4j["Neo4j Graph (agentic_graph_rag)"]
            Qdrant["Qdrant Vector DB (IntelliFlow)"]
            OpenSearch["OpenSearch Lexical Store"]
        end
    end

    subgraph L2["L2: Execution & Event Backbone (Kafka + Temporal)"]
        Kafka["Kafka Event Bus (IntelliFlow)"]
        Temporal["Temporal Workflow Engine"]
        Sandboxes["gVisor / Firecracker Sandboxes"]
        JobRunners["CI/CD & Job Runners"]
    end

    subgraph L1["L1: Core Systems of Record"]
        Postgres["PostgreSQL Core Databases"]
        PerlCore["Perl Legacy Core"]
        TradingSystems["MT5 / cTrader Bridges"]
        CashierCRM["Cashier & CRM APIs"]
    end

    subgraph L0["L0: Cross-Cutting Observability"]
        OTel["OpenTelemetry Spans"]
        Langfuse["Langfuse Tracing & Evals"]
        Prometheus["Prometheus Metrics & Alerts"]
    end

    %% Flow Connections
    UI_Slack & UI_IDE & UI_Web & UI_WA --> PEP1
    PEP1 --> AAT --> Orchestrator
    Orchestrator --> AgentReg & SkillReg & ScopedMem
    Orchestrator --> PEP2 --> PDP

    PDP --> PEP3 & PEP4 & PEP5
    PEP3 --> PII --> Router --> SemCache
    PEP4 --> Broker --> JurisInject --> BlastCap
    PEP5 --> Neo4j & Qdrant & OpenSearch

    BlastCap --> Sandboxes & Temporal & Kafka
    Router --> ModelGW
    Sandboxes & Temporal --> Postgres & PerlCore & TradingSystems & CashierCRM

    %% Control & Observability Enforcement Path
    RiskEngine --> ApprovalSvc --> PDP
    PDP --> CostLedger & AuditLog
    L5 & L4 & L3 & L2 & L1 -.-> OTel & Langfuse & Prometheus
Loading

Pinned Loading

  1. agentic_graph_rag agentic_graph_rag Public

    A production-ready GraphRAG engine using Neo4j, LangGraph, and FastAPI for intelligent retrieval, Text-to-Cypher analytics, and secure graph-native RBAC.

    Python 4

  2. rtldoc rtldoc Public

    Geometry-first PDF parser for RTL/Arabic and complex layouts — reconstructs reading order and bidi text from vector glyph coordinates instead of OCR, with vector-rule table detection and a from-scr…

    Python

  3. uber_lyft_app uber_lyft_app Public

    Uber Lyft app Demo in Flutter inspired by

    Dart 18 16

  4. model-gateway model-gateway Public

    Multi-tenant LLM gateway. One OpenAI-compatible endpoint for every provider, with identity, budgets, policy and PII controls served from an immutable versioned snapshot — so a control-plane outage …

    Go