mudassar hayat

mudassar hayat

Data Scientist & AI Engineer specializing in NLP, LLM Systems, and Agentic AI

Based in: Lahore, PakistanCurrently: Data Scientist, ADDO AIExperience: 4 years

Hire me

Results-driven AI Engineer with 3 years of experience building production-grade machine learning systems, including multi-model AI agents and RAG architectures. He has a proven track record of delivering measurable business impact, such as achieving 90%+ accuracy in automated NLP customer support triaging.

Experience

Data Scientist

ADDO AILahore, Punjabwork2026 – Present

Summary

Worked on a production-grade Al-powered Case Triaging Agent using LangGraph, automating the end-to-end customer support workflow from email classification and information extraction to case investigation and response generation, reducing manual effort for Customer Service Representatives (CSRs)

What I did

  • Fine-tuned and deployed transformer model DistilRoBERTa multiclass email classifier (ETA inquiries, invoice discrepancies, pricing requests, sales inquiries) as a real-time inference endpoint on Aws SageMaker, sustaining low latency.
  • Built an Al information-extraction pipeline (OpenAl, Aws Comprehend, Aws Textract) for email summarization, sentiment/urgency detection, and structured field extraction from invoices, POs, delivery notes, and credit/debit notes documents
  • Extended the document-understanding pipeline to additional business document types and added OcR post-processing/validation logic, improving downstream extraction accuracy.
  • Designed a RAG knowledge-base chatbot (PostgreSQL/pgvector, HNSW indexing, hybrid semantic + BM25/TSVector retrieval) delivering citation-grounded answers with average retrieval latency.
  • Automated case classification, entity extraction from emails and documents, document classification, sentiment analysis, intent classification, and case summarization
  • Implemented initial investigation and final case resolution with human-in-the-loop

Results

  • Reduced manual triage time to 70%
LangGraphDistilRoBERTaAWS SageMakerOpenAIAWS ComprehendAWS TextractOCRRAGPostgreSQLpgvectorHNSWBM25TSVector

Associate Data Scientist

ADDO AILahore, PunjabworkSep 2024 – 2026

Summary

Built and deployed an NLP email-classification system at 90%+ accuracy, automating triage for emails/month and cutting support-team overhead.

What I did

  • Developed a multi-layered document classification pipeline (vision + regex) across images, PDFs, and Excel attachments (Purchase Orders, Invoices, Product Listings, +more), with line-level entity extraction from heterogeneous formats.
  • Implemented automated entity extraction (AWs Comprehend + Regex) to surface PO numbers, invoice IDs, order codes, and product codes without manual review.
  • Built a schema-aware RAG + Text-to-SQL chatbot spanning company policy docs and live database queries, reducing reliance on manual SQL reporting.
  • Transformed robotic surgery logs into interactive visual analytics for medical stakeholders, shortening clinical review turnaround
  • Built a YoLOv7 computer-vision pipeline (annotation → training → inference) to detect trading-signal indicators from financial chart images.
NLPVisionRegexAWS ComprehendRAGText-to-SQLYOLOv7

Associate Data Scientist

LucrumERPLahore, PakistanworkFeb 2024 – Sep 2024

Summary

LucrumERP is ADDO AI's sister company; internally transferred to parent company ADDO AI in Sep 2024.

What I did

  • Built a time-series demand-forecasting model for a retail bakery client, improving forecast accuracy on historical sales data to support inventory decisions.
  • Developed a natural-language sales-database chatbot (LangChain, Text-to-SQL) letting non-technical stakeholders query sales data directly.
  • Built an Al document-intelligence system (LangChain + Ollama/Llama 3) for semantic retrieval and Q&A over DOcX files using ChromaDB vector search.
  • Designed and maintained the data ingestion and preprocessing pipelines supporting the above systems.
  • Developed a SARIMAX-based bakery demand forecasting model to improve prediction accuracy and inventory planning

Results

  • Achieved approximately 75% prediction accuracy across branches
Time-seriesLangChainText-to-SQLOllamaLlama 3ChromaDBSARIMAX

Skills

Technical

PythonPython
LangGraph
YOLO
AWS
SageMaker
LLM Integration
Classification
RAG
Text-to-SQLText-to-SQL
Hugging Face Transformers
Scikit-Learn
Regex
Data Preprocessing
SQLSQL
Vector DBs
GitGit
PyTorchPyTorch
Object Detection
LangChain
OCR
Prompt Engineering
LoRA Fine-Tuning
Ollama
NER
Sentiment Analysis
Feature Engineering
Model Evaluation
Image Segmentation
OpenCVOpenCV
Lambda
EC2
S3
Comprehend
Textract
DockerDocker
MLflow
TensorFlowTensorFlow
FlaskFlask
FastAPIFastAPI
PostgreSQLPostgreSQL
BashBash
Power BI
Model Quantization
Clustering
CodePipeline
KubernetesKubernetes
CI/CD
Azure
GCPGCP
Cross Entropy Loss
SARIMAX
vLLM

Projects

Breast Cancer Image Segmentation U-Net

Lead Developeracademic
Implemented U-Net with and without skip connections for lesion segmentation on medical images, with a custom Dice-loss function, early stopping, and LR scheduling. Evaluated with Dice score: visualized predictions. failure cases, and encoder feature representations via t-SNE
U-NetPyTorcht-SNE

Face Recognition Attendance System

Lead Developeracademic
Built a real-time pipeline (Dlib detection/tracking + VGG16 embeddings) that matched live video against registered profiles, auto-logged attendance, and emailed reports to instructors.
OpenCVDlibVGG16

Image Captioning System

personal
CNN encoder + LSTM decoder, Flickr8K, BLEU evaluation.
CNNLSTMFlickr8K

Medical Image Enhancement

personal
Gaussian/Median filtering, Histogram Equalization, Non-Local Means on noisy X-ray/MRl scans for downstream analysis
Gaussian filteringMedian filteringHistogram EqualizationNon-Local Means

Traffic Vehicle Detection — YOLOv11

Lead Developeracademic
Trained and benchmarked YOLOv11 Nano vs. Medium for vehicle detection, comparing mAP, loU, precision, recall, and inference speed to select the best accuracy/latency trade-off. Built an inference pipeline for detection and visualization on unseen images.
YOLOv11PyTorchUltralytics

Urdu Poetry Classification Research

personal
Conducted research on Urdu poetry classification using a Gemma 2B LoRA fine-tune, comparing performance against traditional ML models, Llama, and BERT baselines.
Gemma 2BLoRALlamaBERT

Education

Master of Science

Information Technology University LahoreData ScienceGPA 3.022022 – 2024

Thesis. Benchmarking Interpretable Multi-Scope & Multi-Level Classification in Urdu Poetry

Coursework

Machine LearningDeep LearningBig Data AnalyticsNLPStatistics for Data ScienceBlockchainMedical Image Computing

Bachelor of Science

University of the PunjabInformation TechnologyGPA 3.212017 – 2021

Thesis. Face Recognition-Based Attendance System using deep learning

Recognition

Certifications

Machine Learning Specialization

Coursera

Deep Learning Specialization

Coursera

Practical Data Science on AWs Cloud Specialization

Coursera

AWS Machine Learning Engineer Nanodegree

Udacity

AWS Cloud Technical Essentials

AWS

Artificial Intelligence Developer

PIAIC

Natural Language Processing

HuggingFace

Computer Vision Course

UET KICS

Patents & publications

Benchmarking Interpretable Multi-Scope & Multi-Level Classification in Urdu Poetry

publication

Paper submitted to TALLIP Journal (currently under peer review). Curated a 43,000+ misra Urdu poetry dataset. Fine-tuned Multilingual BERT and Gemma 2B (via LoRA).

Hire mudassarReach out about a role, a contract or a conversation.For recruiters

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

More people on Kanvis

Browse more peopleCreate your own Kanvis