muhammad hamza

muhammad hamza

AI Engineer specializing in LLM, RAG, and Multi-Agent Systems

Based in: Rawalpindi, PakistanCurrently: AI Engineer, Z Square Solutions (Pvt) Ltd.Experience: 3 years

Hire me

Experienced AI Engineer with 3 years of expertise in building multimodal agentic systems, real-time voice AI, and production-grade RAG pipelines. He has a proven track record of deploying scalable machine learning solutions on AWS using LangChain, LangGraph, and Docker.

Experience

AI Engineer

Z Square Solutions (Pvt) Ltd.Rawalpindi, Pakistanwork2025 – Present

Summary

Built a document retrieval system for natural language search over large unstructured collections, using embeddings and reranking so results match the meaning of a query and not just keyword scores. Also built multi-agent systems where specialized agents work together on complex, multi-step tasks. Built a voice assistant for customer queries and appointment booking, and a medical chatbot for patient triage. Both were designed to handle unclear, noisy user input and give answers people can understand and act on. Worked on a computer vision training platform where users upload images, annotate them, and train detection and segmentation models for smoke detection, medical imaging, and infrastructure monitoring. Ran experiments across model architectures to compare how well each one generalizes. Worked across machine learning, deep learning, NLP, and computer vision in the same production environment, including OCR pipelines for text and number recognition, and evaluated models under real operating conditions.

What I did

  • Designed voice and chatbot systems to handle unclear, noisy user input and give answers people can understand and act on.
  • Built a medical RAG chatbot and document retrieval system using hybrid search over large PDF guideline documents.
  • Evaluated systems using a golden dataset of 150 expert-reviewed questions mapped to source passages and reference answers.
  • Built the Anna and Research Pal production systems.

Results

  • Raised precision@5 from 0.62 to 0.81 by adding a Cohere reranker on top of hybrid search.
  • Reduced unsupported answers by roughly 40% through relevance thresholds and grounding checks.
  • Maintained recall@5 above 0.90 after implementing reranking.
  • Serves around 200 users a day over 5,000+ pages of documents.
  • Achieved a median latency of about 2.5 seconds including the rerank step.
EmbeddingsRerankingMulti-agent systemsVoice assistantMedical chatbotComputer visionDetection modelsSegmentation modelsOCR pipelinesMachine learningDeep learningNLPFastAPILangChainQdrantHybrid SearchCohere RerankerDocker

Junior AI Engineer

Z Square Solutions (Pvt) Ltd.Rawalpindi, PakistanworkAug 2024 – Dec 2024
Built RAG-based and GenAI systems, including client-facing demos that showed stakeholders practical, real use cases. Worked on machine learning problems across different domains, including credit risk modeling, where I compared classification algorithms and evaluated them on real banking data. Promoted to AI Engineer based on project delivery.
RAGGenAIMachine learningCredit risk modelingClassification algorithms

Intern → Software Developer

Soft Wise SolutionsRawalpindi, PakistaninternshipSep 2023 – Jul 2024
Worked with the AI team on object detection, classification, and tracking tasks, exploring YOLO and related Python libraries across different use cases. Experimented with open-source Hugging Face models like Flux and Stable Diffusion for image generation. Explored building RAG-based systems using OpenAI, Claude, and Gemini models together with vector databases, and learned what makes retrieval and generation reliable in practice.
Object detectionYOLOPythonHugging FaceFluxStable DiffusionRAGOpenAIClaudeGeminiVector databases

Skills

Technical

OpenAI
Claude
PythonPython
LangChain
LangGraph
Prompt Engineering
RAG
Agentic AI
GitGit
RAG Evaluation
VAPI
Hugging Face
PyTorchPyTorch
GuardrailsGuardrails
Embeddings
YOLO
ElevenLabs
Multimodal AI
Model Context Protocol (MCP)
n8n
Scikit-Learn
FastAPIFastAPI
Streamlit
Reranking
Data Processing & Cleaning Pipelines
Gemini
Deepgram
OCR
TensorFlowTensorFlow
LlamaIndex
Qdrant
ChromaDB
Monitoring
SQLSQL
PostgreSQLPostgreSQL
FlaskFlask
AWS
CI/CD
LiveKit
DockerDocker
NoSQLNoSQL
C++C++
DVC
Twilio
MLflow
Azure
Testing
Hybrid Search
Cohere Reranker
PostgresSaverPostgresSaver
Schema Validation

Projects

Research Pal

AI Engineerpersonal
Built an AI research assistant using a Retrieval-Augmented Generation (RAG) architecture to extract insights from academic papers. Combined LangChain with a vector database for context-aware retrieval, powering literature review and natural-language Q&A over large document collections. Added a reranking step to improve retrieval quality, along with RAG evaluation, monitoring, guardrails, and a CI/CD pipeline.
RAGLangChainVector databaseRerankingRAG evaluationMonitoringGuardrailsCI/CDLangGraph

Travel Planner

Lead Developerpersonal
Designed a multi-agent trip-planning system where an orchestrator agent interprets a user request (e.g. "plan a 7-day trip to Tokyo") and delegates work to specialized Flight, Hotel, Budget, and Search agents. Used the Model Context Protocol (MCP) so the orchestrator and agents could discover and call the right external tools at runtime instead of relying on hardcoded integrations. An Itinerary agent combines the outputs from all agents into a final day-by-day travel plan. Built with LangGraph for agent state and routing, with PostgreSQL-backed checkpointing for conversation memory. Added reranking, evaluation, monitoring, and guardrails to the agent workflow, with a CI/CD pipeline for deployment.
LangGraphMCPPostgreSQLRerankingEvaluationMonitoringGuardrailsCI/CD

Automated Viral Video Pipeline

Developerpersonal
Built an n8n workflow that turns long YouTube videos into short, ready-to-post clips on its own: it pulls the source video using RapidAPI, splits it into chunks with FFmpeg, and picks out the best short segments. Each clip gets a caption from a vision model (Nano Banana) and a title and tags from an OpenAI model based on the transcript, then is uploaded automatically along with its thumbnail.
n8nRapidAPIFFmpegNano BananaOpenAI

Anna

Developerpersonal
Built an annotation tool and end-to-end ML training platform supporting YOLO, Detectron2, Mask R-CNN, and Faster R-CNN. Users can upload images, annotate datasets, and train custom object detection and segmentation models without external dependencies.
YOLODetectron2Mask R-CNNFaster R-CNNLangGraphComputer Vision

Doojo AI

Developerpersonal
Built a voice-enabled AI assistant using multiple orchestrated agents with tool-calling for tasks like availability checks and appointment scheduling. Combined VAPI, ElevenLabs, Deepgram, and open-source TTS/STT models to keep conversations natural, and tuned the pipeline to keep response latency low so calls felt real-time.
VAPIElevenLabsDeepgramTTSSTT

Credit Risk Modeling Using Banking Data

Student Researcheracademic
Worked through a raw banking dataset using the full machine learning pipeline: data cleaning, feature selection, comparison of classification algorithms in Python, and evaluation. Analyzed why different algorithms gave different results on the same dataset and presented the methodology and findings to faculty.
PythonMachine learning pipelineClassification algorithms

Education

Bachelor of Science

PMAS Arid Agriculture University RawalpindiComputer ScienceOct 2020 – Aug 2024

Thesis. IntelliRoom+, an IoT and computer vision system that automates a whole room

Recognition

Certifications

Develop Generative AI Applications: Get Started

IBM (Coursera)

Azure OpenAI Essentials

Coursera

Machine Learning with Python (V2)

Coursera

Foundations of Project Management

Google (Coursera)
Hire muhammadReach out about a role, a contract or a conversation.For recruiters

muhammad's twin is AI, it can make mistakes.

More people on Kanvis

Browse more peopleCreate your own Kanvis