Jul 2020 – Present

Modernizing Global UN Translation & Document Systems with AI

United Nations · Senior Full Stack Developer

772M+

Documents Indexed

Hybrid

Keyword + Vector Search

6

UN Languages

AI

LangGraph Doc Assistant

Growth & Stability

~60%

API Response Time Reduced

30+

Manual Workflows Eliminated

1000s

Real-time Synced Users

Overview

At the United Nations, I worked as a Senior Full Stack Developer on a suite of mission-critical linguistic and document platforms used globally by translators, editors, diplomats, and researchers. These systems power translation, editing, search, terminology management, and document accessibility across multiple languages and global offices. They run search at internet scale — a production Elasticsearch estate whose largest clusters hold ~772 million and ~100 million documents, plus Azure AI Search across 16 UN duty stations worldwide — with hybrid keyword + vector retrieval feeding a production multilingual RAG assistant — a LangGraph ReAct chatbot (FastAPI + Azure OpenAI) that answers natural-language questions over official UN documents and voting records across all six official languages.

The Problem

Legacy REST APIs limiting flexibility and performance

Massive multilingual datasets across documents and terminology

Fragmented tools for translation, editing, and search

Need for real-time collaboration between translators and editors

Difficulty in making documents machine-readable and searchable

Searching hundreds of millions of multilingual documents with relevance

Limited AI assistance in translation and document retrieval workflows

The Solution

Replaced legacy REST APIs with GraphQL-based architecture

Built AI-powered tooling for translation assistance

Built a production multilingual RAG chatbot (FastAPI + LangGraph ReAct + Azure OpenAI) answering questions over UN documents and voting records

Transcribed recordings of UN speeches into the 6 official languages via AWS Transcribe, orchestrated by Python Azure Functions

Enhanced multilingual search and document retrieval systems

Developed structured document conversion pipelines (Word → XML)

Improved collaboration workflows between translators and editors

Integrated terminology intelligence (UNTERM) into translation flow

Operated Elasticsearch at scale with custom multilingual analyzers and relevance tuning

Built hybrid keyword + vector retrieval feeding a LangGraph ReAct document assistant

Built ETL pipelines into a data warehouse with SSRS reporting, Power BI dashboards, and DAX measures

Architecture

Core Components

eLUNa → translation platformeLUNa Editorial → editing systemeLUNa Search → multilingual search engineUNTERM → global terminology databaseUN Digital Library → public document accessODS → Official Document SystemSpeech-to-text transcription → 6 UN languages (AWS Transcribe)AI Assistant → multilingual RAG chatbot (UN documents + voting records)

Technical Design

Backend: C# / .NET services (incl. authored/modified legacy internal WCF SOAP services)
API Layer: REST → GraphQL migration
Data: SQL Server + XML-based document structures
Data pipelines: ETL into a data warehouse, with SSRS reporting and Power BI dashboards / DAX measures
Search: multiple production Elasticsearch clusters (largest ~772M docs across 2,106 indexes / ~800 GB, another ~100M) plus Azure AI Search across 16 UN duty stations
Retrieval: hybrid keyword + vector search with custom multilingual analyzers
Frontend: Vue.js applications
Cloud: Azure, with VNET / private endpoints / VPN for private service-to-service connectivity
AI backend: Python / FastAPI (uvicorn + gunicorn), LangChain + LangGraph ReAct agent, Azure OpenAI (GPT-4o-mini), Managed Identity auth (no API keys)
RAG retrieval: Azure AI Search (HNSW ANN + semantic hybrid — BM25 + vector fused, then a semantic reranker), with per-language indexes across the 6 official languages
Multi-agent chatbot with a voting-data SQL agent (LangChain SQLDatabaseToolkit, two-stage schema-safety design)
Anti-hallucination guardrails (verbatim quotes + per-claim citations, refuse-on-no-retrieval, reranker thresholds, negative-dataset evals); OpenAI-compatible streaming API; customized LibreChat UI
Speech-to-text: Python Azure Functions calling AWS Transcribe (6-language), secrets in Azure Key Vault (cross-cloud Azure + AWS)

My Role: Senior Full Stack Developer

01

Designed and implemented GraphQL APIs replacing legacy REST systems

02

Built and enhanced features across translation, editing, and search platforms

03

Developed document conversion pipeline (Word → structured XML / AKN4UN)

04

Improved performance and scalability of large document systems

05

Integrated AI capabilities into existing platforms

06

Built cross-cloud speech transcription: Python Azure Functions calling AWS Transcribe to transcribe UN speeches into 6 languages

07

Operated production Elasticsearch at scale (7 clusters, ~772M documents in the largest)

08

Built hybrid keyword + vector search and a LangGraph ReAct document assistant

09

Built the production multilingual RAG chatbot (FastAPI, LangGraph, Azure OpenAI, Azure AI Search) over UN documents and voting data

010

Built ETL pipelines into a data warehouse feeding SSRS reports and Power BI dashboards (DAX)

011

Collaborated with product managers, UX teams, and domain experts

012

Contributed to frontend systems using Vue.js

013

Wrote automated tests (unit, integration, Selenium)

014

Supported DevOps workflows on Azure

Key Achievements

🔥

Migrated multiple systems from REST to GraphQL architecture

🔥

Built AI-powered translation assistant used in production workflows

🔥

Enabled machine-readable UN documents through structured XML conversion

🔥

Operated Elasticsearch at internet scale — ~772M documents in the largest cluster

🔥

Delivered hybrid keyword + vector retrieval feeding a LangGraph document AI assistant

🔥

Shipped a production RAG chatbot answering natural-language questions over UN documents and voting records in 6 languages

🔥

Met SLA-backed 99.9% uptime for the governments and inter-governmental organizations relying on the platforms

🔥

Enhanced real-time collaboration between translators and editors

🔥

Contributed across multiple mission-critical UN platforms

Impact

📊

Improved efficiency of translators through automation and AI assistance

📊

Enabled better access to global UN documents for public and internal users

📊

Increased system scalability and maintainability through API modernization

📊

Reduced complexity in querying large multilingual datasets

📊

Enhanced consistency and accuracy of translations using terminology systems

📊

Supported global users working across multiple languages and regions

📊

Sustained SLA agreements and 99.9% uptime with member states, inter-governmental, and government organizations

Tech Stack

Backend

C#, .NET, GraphQL, REST, WCF/SOAP (legacy services)

Frontend

Vue.js

Cloud

Azure

Data

SQL Server, XML

Data Eng

SSRS, ETL, Data Warehouse, Power BI, DAX

Search

Elasticsearch (7 clusters, ~772M docs), Azure AI Search, hybrid keyword+vector

AI

Python, FastAPI, LangChain, LangGraph (ReAct), Azure OpenAI (GPT-4o-mini), Azure AI Search (semantic hybrid), Azure AI Foundry, Managed Identity, MongoDB

Chatbot

Multilingual RAG, multi-agent tools, voting-data SQL agent, LibreChat UI, OpenAI-compatible streaming

Speech

AWS Transcribe (6-language), Azure Functions, Azure Key Vault

Networking

VNET, private endpoints, VPN

Testing

xUnit, Selenium, Integration Testing

DevOps

Azure DevOps, Terraform, Azure Bicep (IaC)

Observability

Application Insights (consolidated from Seq)

What I Help With

Scalable Backend Systems

High-performance APIs and microservices

Complex Integrations

Stripe, Shopify, Microsoft Graph, telecom

AI & Automation

LLM workflows, document processing

Identity & Security

OAuth, Entra ID, Cognito

Let's discuss your system or project

Get expert input on your architecture, integrations, or scaling challenges.

Book a Call to Discuss Your Project