Haris Majeed

AI Engineer specializing in Computer Vision pipelines and Multimodal Deepfake Detection.

Based in: Lahore, PakistanCurrently: AI Engineer, MLBench Pvt LtdExperience: 2 years

Haris is an AI Engineer at MLBench with a strong background in building production-grade computer vision systems and full-stack RAG applications. He holds a BS in Computer Science and has successfully developed novel deepfake detection pipelines and optimized document extraction models.

Experience

AI Engineer

MLBench Pvt LtdworkJun 2026 – Present

Summary

Worked on a field detection system for document forms, replacing a client's three-model pipeline (YOLO, FFDNet/FFDetr, and a geometry classifier) with a single fine-tuned RF-DETR model. Fine-tuned RF-DETR-Base on an annotated document dataset, reaching 0.9335 precision and 0.8496 recall against the client's existing three-model architecture at 0.9394 precision and 0.7875 recall, an F1 improvement from 0.8568 to 0.8896. Evaluated Tesseract and DocTR as OCR backends for the extraction pipeline, identifying checkbox glyph misreads and low-resolution footer text corruption in Tesseract output that DocTR resolved. Annotated client PDF documents for field geometry, type, and role to build the training set for the detection model, and logged experiment results to guide model selection. Contributed to a rice grain quality and variety analysis tool, training DINOv2-based classifiers across multiple image resolutions to grade grains as Full, Broken, Rejected, Weak, or Fatty and identify variety among Full grains.

What I did

Provided an API endpoint for the Inspectech project to the web team for frontend integration.

Results

Improved F1 score from 0.8568 to 0.8896 for document field detection.
Successfully replaced a three-model pipeline with a single fine-tuned RF-DETR model.

YOLOFFDNetFFDetrRF-DETRRF-DETR-BaseTesseractDocTROCRDINOv2Hugging Face

Teaching Assistant - Software Engineering & Artificial Intelligence

Information Technology UniversityLahorepart timeAug 2025 – Jun 2026

Summary

Assisted in teaching core Software Engineering and Artificial Intelligence concepts, including system design, version control, testing methodologies, search algorithms, and fundamentals of machine learning. Mentored students in developing scalable software and AI projects, guiding model selection, experiment design, and evaluation. Conducted code reviews, graded assignments and exams, and provided constructive feedback to strengthen student learning. Provided technical guidance on Python and ML frameworks including scikit-learn and PyTorch, and facilitated discussions linking course material with current AI research trends.

What I did

Provided additional mentorship outside of office hours to help students identify real-world problems and optimal technical solutions.

Results

Managed around 120 students and almost 60 project groups simultaneously.

Pythonscikit-learnPyTorchVersion Control

Front-End Intern

CerturainternshipJul 2025 – Aug 2025

Built a responsive portfolio website using HTML, CSS, and JavaScript with animations and media query-based responsiveness. Developed a dynamic to-do list application and a weather dashboard using the OpenWeather API, with real-time data fetching and error handling.

HTMLCSSJavaScriptOpenWeather API

Skills

Technical

Python

VS Code

RF-DETR

Git

C++

Jupyter

scikit-learn

PyTorch

YOLO

pandas

NumPy

RAG Pipelines

Groq

CLIP

BLIP-2

HuggingFace Transformers

FastAPI

SQL

HTML

CSS

JavaScript

React

Flask

MongoDB

Matplotlib

LangChain

FAISS

DINOv2

ByteTrack

OpenCV

Tesseract

DocTR

JWT Authentication

Linux (Ubuntu)

Salesforce

Node.js

Qwen-VL

Gradio

Java

Cellpose-SAM

Express

Apex

Logisim

Recursive-descent parsing

Hugging Face

Dead code elimination

Projects

Multimodal Deepfake Detection using VLMs (FYP)

Lead Researcher/Developeracademic

Benchmarked state-of-the-art deepfake detectors (SBI, AltFreezing, Face Warping Detection) across FaceForensics++, Celeb-DF-v2, and HiDF, revealing that binary classifiers collapse to near-chance accuracy on high-quality deepfakes and have no mechanism to explain their predictions. Designed ExDDV-Judge, a novel three-stage detect, explain, judge pipeline that replaces 11B/7B-parameter VLMs with lightweight alternatives (BLIP-2 T5-XL, LLaVA-Mini), cutting model size by over 80% while running entirely on a single consumer GPU. Solved a critical real-frame misclassification bug in prior work by building a Dual-DB routing system with a fine-tuned CLIP encoder, taking real-frame accuracy from a broken 0% to 83%. Connected explanation generation with automated verification for the first time by pairing BLIP-2 T5-XL with DeepfakeJudge, producing faithfulness-scored explanations. Shipped a full-stack Flask web app delivering real-time, forensic-grade deepfake analysis, including per-frame labels, natural-language explanations, and trust scores, turning a research pipeline into a usable end-to-end tool.

SBIAltFreezingFace Warping DetectionBLIP-2 T5-XLLLaVA-MiniCLIPDeepfakeJudgeFlaskPython

Legal AI Chatbot

Full Stack Developerpersonal

Built a full-stack RAG-based legal assistant using React, FastAPI, and MongoDB, with document-grounded Q&A powered by LangChain, FAISS, and the Groq API (Llama 3.3 70B). Implemented project-scoped retrieval so each client/case's uploaded PDFs are chunked, embedded, and queried in isolation, with every answer citing its source document. Added JWT-based authentication, persistent conversation memory, and real-time token streaming for a production-style chat experience. Designed the system to run GPU-free by offloading LLM inference to Groq's hosted API, keeping only lightweight CPU-based embeddings local.

ReactFastAPIMongoDBLangChainFAISSGroq APILlama 3.3 70BJWT

View the code

Footfall Counter

Developerpersonal

Built a real-time people-counting system using YOLO/RT-DETR for detection and ByteTrack for persistent multi-object tracking across occlusion. Designed directional entry/exit counting via an arbitrary-angle virtual line, with an interactive tool to configure the line to match any camera setup. Optimized the full pipeline to run entirely on CPU, with configurable confidence, IoU, and resolution parameters for accuracy/speed trade-offs.

YOLORT-DETRByteTrackPythonOpenCV

View the code

Airbnb Clone (Extended)

Full Stack Developerpersonal

Developed a full-stack Airbnb-inspired web application using React, Node.js, and MongoDB. Implemented JWT-based authentication, role-based access control, and protected admin routes. Built features for property listing, booking management, and host-specific CRUD operations.

ReactNode.jsMongoDBJWT

View the code

Search Engine

Developeracademic

Built a custom search engine in C++ using an inverted index for efficient text retrieval. Implemented multi-condition query handling and result ranking. Deployed locally with a user-friendly command-line interface for querying.

C++

View the code

Client-Server Network Communication Project

personal

Developed a pipeline in C allowing multiple clients to communicate and perform file read/write operations over a network without blocking.

C++ Compiler From Scratch

Developeracademic

Built a toy C++ compiler from scratch in C++ to explore core compiler design concepts. Implemented lexical analysis to tokenize input source code. Developed a recursive-descent parser for syntax analysis and basic semantic checks. Handled error detection and parsing of grammar defined in a custom specification. Explored the foundations of compiler pipelines, including future support areas such as code generation.

C++recursive-descent parser

View the code

Inspectech

personal

An AI project initially deployed on Hugging Face for direct user interaction before transitioning to an API-based backend for a dedicated web frontend.

Hugging Face

Education

Bachelor of Science

Information Technology UniversityComputer ScienceGPA 3.28Aug 2022 – Jun 2026

Thesis. Multimodal Deepfake Detection using VLMs

Coursework

Software EngineeringArtificial IntelligenceMachine LearningCompiler Design

Recognition

Certifications

Salesforce Administrator Training

Trailhead2024

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

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

More people on Kanvis

Browse more people Create your own Kanvis