Based in Karachi, Pakistan  ·  Open to Relocation Globally

Ali
Zain

Artificial Intelligence Engineer

Building intelligent systems at the frontier of RAG, agentic AI,
and multilingual NLP — from research to production.

View My Work Get In Touch
3
ACL Publications
#1
M-DAIGT / Urdu Rank
3+
Years Building AI
Scroll
About Me

The Mind
Behind the Models

I'm an AI Engineer with deep hands-on experience in agentic systems, retrieval-augmented generation (RAG), and vertical AI agents. My work lives at the intersection of cutting-edge research and production-grade engineering — I ship systems that actually work.

"Recognized internationally with top rankings in AI-generated text detection benchmarks across M-DAIGT, AraGenEval, and AbjadNLP."

I specialize in taking ambitious AI systems from prototype to production — measurably improving agent performance, building multilingual transformers that top leaderboards in Urdu and Arabic NLP, and shipping infrastructure that scales. I'm currently seeking a full-time AI Engineering role and am open to relocation globally.

// Contact Info
Karachi, Pakistan
vin.alizain@gmail.com
linkedin.com/in/alizain-157
github.com/alizain-157
RAG Systems Agentic AI NLP Research LLM Fine-tuning Open to Relocation
Technical Skills

The Full
Stack

🤖
// AI / LLM Systems
RAG Pipelines Multi-Agentic Systems Prompt Engineering Fine-tuning Reinforcement Learning Evaluation OpenAI Gemini Claude
🧠
// Frameworks
PyTorch TensorFlow HuggingFace LangChain LangGraph CrewAI AutoGen Agno FastAPI
💾
// Databases
Pinecone Weaviate Qdrant MongoDB PostgreSQL Redis
☁️
// Cloud & Infra
AWS AWS SageMaker AWS Bedrock Github Actions DevOps
💻
// Languages
Python C/C++ SQL / NoSQL JavaScript / TypeScript HTML / CSS
🛠️
// Tools & Other
Weights & Biases N8n OpenCV Next.js React.js Keras
Work Experience

Where I've
Shipped

Dec 2025 — Present
San Francisco, CA (Remote)
Traversaal.ai
Artificial Intelligence Engineer
  • Collaborated on vertical AI agents for Data Science and Machine Learning tasks; rigorously tested the Data Science Agent across CSV and SQL datasets, validating Python/SQL code generation and visualizations.
  • Improved Data Science Agent performance on DSBench leaderboard from 59% → 77.5% by implementing a debugger loop and context engineering techniques.
  • Enhanced RAG pipelines by decomposing complex user queries into atomic sub-questions, resulting in significantly better recall and answer quality.
Jul 2025 — Dec 2025
San Francisco, CA
QLU.ai
Applied AI Research Engineer
  • Advancing QLU.ai's mission of transforming executive hiring through multimodal AI by researching and engineering text, speech, and video models.
  • Developed a TTS and voice cloning system that improved timbre quality and naturalness, enabling context-aware and personalized voice interactions.
  • Building Computer Use Agents capable of autonomously operating graphical interfaces, leveraging data acquisition pipelines and Reinforcement Learning (RL).
  • Initiated development of a high-precision design system targeting intelligent candidate retrieval on complex queries within 5–10 seconds.
Aug 2024 — Mar 2025
San Francisco, CA (Remote)
Ragioneer
Full Stack AI Engineer
  • Designed and deployed RAG pipelines powering large-scale recommendation and reasoning systems with improved scalability and accuracy.
  • Collaborated on end-to-end development of GlobalGate, an AI-driven visa recommendation platform, handling frontend (Next.js) and backend-AI architecture (FastAPI, LangChain).
  • Optimized multi-model inference pipelines and caching strategies, reducing latency from 300s → 25s and boosting recommendation accuracy by 15%.
Projects

Things I've
Built

// PROJECT 01
AI vs. Human Text Detection — Black Box Approach
A robust binary classification system capable of distinguishing AI-generated text from human-written content across multiple domains. Evaluated classical and neural strategies over large-scale datasets (700K train / 260K test). Created "Candace" — a LLaMA-3.2–based encoder achieving 88.1% accuracy, outperforming larger models in efficiency.
RoBERTa T5 LLaMA-3.2 XGBoost TF-IDF PyTorch
// PROJECT 02
Cloudy.ai — Speech-to-Speech AWS Cloud Assistant
A speech-to-speech AI assistant that helps users navigate AWS Cloud services in real-time. Integrated Gemini multimodal reasoning with STT (AssemblyAI) and TTS (ElevenLabs). Built low-latency pipelines with Redis queues and async FastAPI for seamless real-time interaction.
Gemini FastAPI Redis MongoDB React.js ElevenLabs
// PROJECT 03
GlobalGate — AI-Driven Visa Recommendation Platform
End-to-end AI-powered visa eligibility guidance platform. Handled full-stack development from frontend to backend-AI architecture. Multi-model RAG pipelines delivered real-time eligibility guidance at scale with dramatically reduced latency.
Next.js FastAPI LangChain RAG Pinecone
// PROJECT 04
Advanced Customer Support Agent — A2A + MCP + ADK
A production-grade, multi-agent customer support system built on Google's Agent Development Kit (ADK), the Agent2Agent (A2A) protocol, and Model Context Protocol (MCP). Specialized sub-agents communicate peer-to-peer via A2A, while MCP standardizes tool and data-source access across the system. Enables autonomous ticket triage, contextual resolution, and seamless escalation through a fully orchestrated multi-agent pipeline.
Google ADK A2A Protocol MCP Multi-Agent Systems Python
Publications & Achievements

Peer-Reviewed
Research

EACL 2026
LoRAD: Low-Resource AI-Generated Text Detection with XLM-RoBERTa
Fine-tuned XLM-RoBERTa for binary classification of human vs. machine-generated text in Arabic and Urdu. Leveraged raw text preservation for Urdu and data augmentation for Arabic, demonstrating the effectiveness of multilingual transformers for low-resource AI-text detection. Published in AbjadNLP 2026, co-located with EACL 2026.
🥇 1st — Urdu  |  🥉 3rd — Arabic
EMNLP 2025
A Comparative Study of Transformer-Based Models for Arabic AI-Generated Text Detection
Evaluated AraELECTRA, CAMeLBERT, and XLM-RoBERTa for binary classification of human vs. AI-generated Arabic text. XLM-RoBERTa achieved the best F1-score (0.7701), surpassing specialized Arabic models and demonstrating strong cross-lingual generalization. Published in EMNLP 2025 Proceedings (ACL).
🏅 5th Overall
RANLP 2025
A Multi-Strategy Approach for AI-Generated Text Detection
Developed three AI-generated text detection systems for the M-DAIGT Shared Task: a fine-tuned RoBERTa classifier, a TF-IDF + SVM model, and the Candace ensemble system. Proposed a novel LLaMA-Feature Ensemble using probabilistic indicators from multiple LLaMA-3.2 models. Achieved near-perfect F1 = 99.9% on both news and academic datasets. Published in RANLP 2025 Proceedings (ACL).
⚡ F1: 99.9%
Education

Academic
Foundation

Bachelor of Science in Computer Science
National University of Computer and Emerging Sciences (FAST-NUCES)
📍 Karachi, Pakistan  ·  Sep 2021 – Jun 2025
3.24
CGPA
// Relevant Coursework
Information Retrieval Deep Learning Natural Language Processing DevOps Generative AI
Let's Connect

Ready to
build together?

I'm currently open to full-time AI Engineering roles globally. If you're working on something ambitious, let's talk.

vin.alizain@gmail.com linkedin.com/in/alizain-157 github.com/alizain-157 +92 326 359 0092