Wajahat Hassan

Machine Learning Engineer Specializing in Agentic AI and Multi-Agent Orchestration

Based in: Islamabad, PakistanCurrently: Machine Learning Engineer, Red BufferExperience: 4 years

A computer science graduate from NUST with extensive experience in building production-grade AI pipelines, ranging from GeoGNN-based connectivity modeling to multi-agent orchestration using LangGraph. He possesses a deep technical background in Computer Vision, NLP, and Data Engineering, complemented by a proven track record of delivering scalable solutions in medical imaging and automated lead generation.

Experience

Machine Learning Engineer

Red BufferworkJul 2025 – Present

Summary

Built a production Survey-to-CAD automation pipeline transforming raw survey archives into AI-ready graph data, standardizing 1,462 historical projects into a unified Intermediate CAD Specification (ICS) for consistent model training. Designed a hybrid feature-code mapping strategy to handle real-world naming variance, achieving 98.3% mapping coverage and reducing unmapped codes by 57.2%; implemented project-level quality controls keeping unmapped rates to 0-4.3% per project. Developed and iteratively improved a GeoGNN-based connectivity modeling pipeline with architecture scaling, semantic feature fusion, and threshold tuning - delivering a strong all-code validation baseline (F1: 0.701, Precision: 0.638, Recall: 0.778, Accuracy: 0.928). Built an AI-based lead generation platform that transformed one company input into a complete outbound package: market research, ideal customer profile definition, target company discovery, contact identification, and contact verification. - using LangGraph and OpenAI agents with Apollo API enrichment. Implemented contact quality validation with ZeroBounce, SearXNG, and LLM-based checks to clean email and phone data before outreach. Shipped with a Streamlit frontend, FastAPI microservices, CSV/JSON/PDF exports, and a Docker Compose stack (PostgreSQL, Redis, Elasticsearch). Built a platform linking contracts to their full negotiation history via unified Gmail and Outlook synchronization (OAuth 2.0, batch ingestion), with OpenAI embedding pipelines and PostgreSQL pgvector for semantic search across email threads. Engineered a three-stage cascade clause analysis pipeline (semantic search → LLM filtering → final analysis) with multi-agent orchestration using the OpenAI Agents SDK, reducing LLM API calls by 60%. Designed specialized sub-agents for email and contract retrieval, coordinated by a central synthesis agent; architected an async FastAPI backend with Redis caching, S3 storage, and JWT-based auth with database-level user isolation. Developed a CLI and web-based incident management system with session-based memory for contextual continuity across multi-step investigations, and Model Context Protocol (MCP) integration for incident retrieval, root cause analysis, and action plan generation. Engineered multi-agent orchestration with LangGraph using concurrent and sequential execution phases for data analysis, similar-incident retrieval, and synthesis; delivered via a Flask web server with an HTML/CSS/JS frontend for interactive reporting.

What I did

Developed a contract analysis platform using a RAG-based architecture to retrieve information from contracts and email conversations.
Built a three-stage pipeline involving retrieval, LLM filtering, and final analysis.
Optimized a GeoGNN-based model for predicting connectivity between survey features by experimenting with architectures, feature combinations, and prediction thresholds.
Improved data quality and feature representation to address inconsistencies in survey data.

Results

Standardized 1,462 historical projects into a unified Intermediate CAD Specification (ICS).
Achieved 98.3% mapping coverage and reduced unmapped codes by 57.2%.
Delivered GeoGNN validation baseline: F1: 0.701, Precision: 0.638, Recall: 0.778, Accuracy: 0.928.
Reduced LLM API calls by 60% through a three-stage cascade clause analysis pipeline.

Graph Neural NetworksGeoGNNLangGraphOpenAI Agents SDKApollo APIZeroBounceSearXNGStreamlitFastAPIDocker ComposePostgreSQLRedisElasticsearchOAuth 2.0pgvectorS3JWTModel Context Protocol (MCP)FlaskHTMLCSSJavaScriptOpenAI embeddings

Machine Learning Engineer

Data QuartzworkNov 2024 – Jun 2025

Summary

Employed gladia, talkscriber and whisper to develop a real-time accent modification tool. Managed data orchestration pipelines using Prefect and Airflow to map customer calls with optimal agents based on historical data and satisfaction metrics. Automated ETL processes for real-time data integration from multiple sources, ensuring efficient and accurate customer-agent matching. Enhanced operational efficiency by automating data scraping, cleaning, and model training through CRON job configuration, resulting in improved data processing time. Engineered a computer vision-based suspicious behavior detection system using YOLOv8 pose estimation and LSTM networks with attention mechanisms, implementing person tracking algorithms and a complete pipeline from data labeling to real-time inference with visualization capabilities.

What I did

Experimented with sentence confidence intervals to ensure smoothness in the accent modification system.
Developed a fracture detection system using a segmentation-based approach and various backbones before selecting YOLO.

Results

Achieved around 95 percent accuracy for the fracture detection system using YOLO and custom datasets.
Reduced latency to around 0.7 seconds for the real-time accent modification system.

gladiatalkscriberwhisperPrefectAirflowCRONYOLOv8LSTMAttention MechanismsYOLO

Machine Learning Intern

xFlow Research Inc.internshipAug 2024 – Oct 2024

Summary

Engineered a custom fracture detection pipeline using YOLOv8, optimizing for X-ray imaging data with targeted class filtering and augmentation. Streamlined preprocessing of 5,000+ medical images, automating annotation refinement and label filtering for fracture-specific object detection. Fine-tuned YOLOv8 model over 80 epochs, leveraging transfer learning to improve localization accuracy and detection confidence. Optimized model hyperparameters and applied advanced techniques like multi-scale training and mosaic augmentation to enhance detection robustness. Achieved high mAP (mean average precision) on test data, delivering state-of-the-art results in fracture identification with minimal false positives. Deployed model evaluation pipeline with custom post-processing scripts for validation on unseen medical datasets.

Results

Processed 5,000+ medical images.
Achieved high mAP on test data.

YOLOv8X-ray imagingTransfer LearningMulti-scale trainingMosaic augmentation

Skills

Technical

Python

Computer Vision

NLP

LangGraph

OpenAI Agents SDK

LangChain

YOLOv8

Generative AI (LLM)

Agentic AI

FastAPI

Multi-Agent Orchestration

TensorFlow

Graph Neural Networks

Data Engineering

Web Scraping

PyTorch

Keras

scikit-learn

Transformers

Flask

Streamlit

PostgreSQL

pgvector

Docker

Git

Selenium

SQL

RAG

AWS

SQL Server

ViT

spaCy

Redis

Medical Imaging

DuckDB

Airflow

Prefect

Oracle

AWS S3

Pydantic

Structured Decoding

Cosine similarity

Unsloth

Quantization

LiteLLM

Surprise library

Areas of expertise

Leadership

Management

Interpersonal Skills

Education

Bachelors in Computer Science

National University of Sciences and TechnologyComputer ScienceSep 2019 – Jun 2024

Societies & activities

NUST Cricket Team (Member, Captain)

Recognition

Awards

Victory in inter-NUST tournament

National University of Sciences and Technology

Cricket tournament victory

Social impact & activities

Volunteer

Rizqvolunteer2021 – 2024

Volunteered during Ramadan, organizing food drives to support underprivileged communities. Assisted in setting up communal tables and ensuring smooth distribution of meals and refreshments during Iftar.

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

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

More people on Kanvis

Browse more people Create your own Kanvis