Rumman J.

Key Skills

Software

Label Studio

Argilla

CVAT

Encord

SuperAnnotate

Snorkel AI

Prodigy

Top Subject Matter

Customer support intent classification and multilingual localization for routed responses.

NLP emotion recognition classification.

Deep learning model optimization (quantization, pruning, knowledge distillation) and evaluation.

Top Data Types

Text

Computer Code Programming

3D Sensor

Top Task Types

Fine-tuning

RLHF

Computer Programming/Coding

Classification

Text Summarization

Question Answering

Text Generation

Object Detection

Segmentation

Freelancer Overview

Customer Support Intent Classification Gateway (E-Commerce AI Routing Hub). Brings 5+ years of professional experience across complex professional workflows, research, and quality-focused execution. Core strengths include HuggingFace Transformers and HuggingFace. Education includes Master of Science, EPITA – École Pour l'Informatique et les Techniques Avancées (2025) and Bachelor of Science, COMSATS University Islamabad, Lahore Campus (2020). AI-training focus includes data types such as Computer Code, Programming, and Text and labeling workflows including Fine-tuning, Evaluation, and Rating.

Labeling Experience

AI Model Optimization Platform (Master's Thesis)

Developed an AI model optimization workflow that ranks and iteratively improves deep-learning models using compression techniques. Applied quantization, pruning, and knowledge distillation to reduce model size while controlling accuracy degradation. Integrated interactive tooling to recommend optimizations and monitor sustainability metrics during evaluation. • Ranked 1/106 team projects with thesis-validated optimization results. • Achieved up to 80% model-size reduction with ~10% inference-latency improvement. • Kept accuracy loss below 5% across evaluated ResNet architectures. • Added real-time carbon-footprint tracking and a Groq-powered LLaMA assistant for guidance.

2025

Sentiment Analysis — NLP Architecture Comparison

TextFine-tuning

Trained and benchmarked multiple tweet-emotion recognition architectures on a 6-class emotion dataset with strict train/validation/test splits. Compared an FCNN baseline, LSTM/GRU sequence models, and a fine-tuned HuggingFace Transformer. Produced an analysis of trade-offs between model complexity, training time, and classification performance. • Implemented rigorous data splitting to prevent data leakage across 6 emotion classes. • Fine-tuned a HuggingFace Transformer to obtain the best accuracy and F1. • Benchmarked FCNN and LSTM/GRU against Transformer performance. • Documented comparative results and model trade-offs for NLP model selection.

2025

Customer Support Intent Classification Gateway (E-Commerce AI Routing Hub)

Fine-tuning

Built and evaluated an NLP intent classification model by fine-tuning a DistilBERT sequence classifier on customer-support dialogue data. Implemented a routing engine that dispatches deterministic responses for high-confidence predictions and triggers human-in-the-loop escalation for ambiguous cases. Added multilingual translation components to enable processing of French tickets using an English-trained model. • Fine-tuned DistilBERT for intent labels: WHERE_IS_MY_ORDER, REFUND_REQUEST, CANCEL_ORDER, PRODUCT_FEEDBACK. • Achieved 96.8% evaluation accuracy with confidence-threshold routing (72%). • Used MarianMT + SentencePiece for language detection and bidirectional translation. • Developed FastAPI confidence-threshold logic and Human-in-the-Loop escalation workflow.

2025

Education

Work History