afftab.

github linkedin x

I study human cognition to code better.

# work.tex

Georgia State University
2025 - 2026 Master’s of Science in Computer Science
  • 2025 - 2026 Master’s of Science in Computer Science, Georgia State University, Atlanta, GA
Ecommerce.co (Remote)
Senior Software Engineer 2025
  • Developed ecommerce marketplace SaaS using Typescript, Next.js, MongoDB, and GraphQL.
  • Built data pipelines for Shopify workflow integration and management.
  • Optimized performance to scale to 10,000 active users/month.
  • Developed AI Agent for natural language product search and keyword extraction.
WinningHunter (Remote)
Software Engineer 2023 - 2026
  • Developed SaaS platform features using PHP and JavaScript, RESTful APIs, responsive UI components.
  • Implemented AI-driven features focusing on UX and cross-functional stakeholder communication.
  • Built cron jobs and ETL processes for automated data ingestion and database maintenance.
  • Scaled platform to 50,000 monthly users, $2M+ annual revenue through performance optimization.
  • Improved platform reliability through QA, unit testing, and iterative product improvements.
TReNDS Center
Graduate Research Assistant May 2026
  • Conducting neuroimaging and data science research at Center for Translational Research in Neuroimaging and Data Science.
  • Developing mindfultensors PyPI package for tensor operations and deep learning optimizations.
  • Contributing to wirehead distributed caching system for synthetic data generation in ML pipelines.
  • Researching fMRI analysis, holographic training methods, and neuroimaging optimization techniques.
Georgia State University
Graduate Research Assistant 2025 - 2026
  • Developed multimodal AI clinical decision support system using LLMs for medical triage scenarios.
  • Applied RAG and fine-tuning with GPT via OpenAI API, achieving 95% F1 accuracy on medical NLP tasks.
  • Explored multimodal vision, speech, and text interaction; evaluated model performance using precision, recall, F1 metrics.
Southeast Missouri State University
Graduate Research Assistant 2024
  • Curated open-source dataset of 400 peer reviews including data cleaning, annotation, and documentation.
  • Fine-tuned LLMs (DistilBERT, RoBERTa, XLNet) for NLP peer-review text classification using supervised learning.
  • Achieved 76.71% validation accuracy with XLNet through loss function optimization and hyperparameter tuning.
  • Evaluated model performance using accuracy, precision, recall, F1 metrics; compared BERT vs GPT models.
Projects
Gradient Quantization for Federated Text Classification Feb 2026 - Apr 2026
  • Implemented adaptive gradient quantization for federated learning in text classification tasks to reduce communication overhead across edge devices.
  • Developed dynamic bit-width quantization (2-8 bits) for gradient compression in NLP models using PySyft and FedAvg algorithm.
  • Evaluated on sentiment analysis with DistilBERT across multiple client simulations, achieving 90%+ of full-precision accuracy.
  • Demonstrated significant communication reduction (8-16x bandwidth savings) while maintaining model performance in federated NLP settings.
Quantization-Aware Fuzzy Calibration for Edge LLMs Jan 2026 - Apr 2026
  • Developed novel fuzzy-gated Dirichlet calibration method for interpretable uncertainty quantification in transformer text classification.
  • Implemented cross-architecture calibration across 4 models (FinBERT, FinancialBERT, Gemma 3, Qwen 3.5) achieving 59-89% Expected Calibration Error (ECE) reduction.
  • Discovered counterintuitive finding that general financial training outperforms narrow task specialization by 5.56% accuracy.
  • Achieved optimal results with FinBERT: 60.08% accuracy, 0.033 ECE, 132 MB memory footprint suitable for edge deployment.
  • Paper in preparation for IEEE Machine Learning for Signal Processing (MLSP) 2026 conference.
Zero-Shot Classification for On-Device LLM Benchmarking
  • Developed an intelligent prompt routing system using zero-shot classification with a lightweight TinyLlama-1.1B-Chat model to select optimized quantized LLMs for each prompt category.
  • Implemented multi-model routing between specialized Qwen3-0.6B quantized models to enhance classification accuracy, efficiency, and generation quality across factual, reasoning, creative, instruction-heavy, and role-based tasks.
  • Designed a comprehensive benchmarking framework with curated prompts and performance metrics (energy consumption, memory usage, latency, throughput, accuracy).
  • Achieved more than 60% classification accuracy and demonstrated 15-30% performance improvement using auto-routing over single-model baselines, enabling efficient on-device execution with significant resource savings.
Vision Studio
  • Created a no-code computer vision platform with React/TypeScript for dataset labeling and model fine-tuning.
  • Enabled fine-tuning of ImageNet models using TensorFlow and automated annotation with YOLOv8.
  • Developed Django backend with PostgreSQL for efficient storage and retrieval.
mindfultensors
  • Developed and published PyPI package for tensor operations and deep learning workflow optimizations.
  • Implemented efficient tensor manipulation utilities and memory optimization techniques for large-scale ML applications.
  • Designed modular architecture supporting multiple frameworks (PyTorch, TensorFlow, JAX).
  • Active maintenance and documentation development for open-source community.
Research Contribution
Published Work 2025

Aashir Aftab, Junaid Shuja et al. “Classifying Scientific Peer Reviews: Distinguishing Authentic, Generic, and AI-Generated Feedback.” Proc. 2025 IEEE International Conference on Future Machine Learning and Data Science (FMLDS 2025), Los Angeles, CA, USA, Nov. 2025

Published Work 2025

Aashir Aftab, Eyal Aharoni, Ttanvi Tummapudi. “AI-Assisted Medical Triage Training” Poster presented at the 2025 Innovations in Artificial Intelligence (AI) Conference, Little Rock, AR, USA, Oct. 2025.

Published Work 2026

Eyal Aharoni, Aashir Aftab, Ttanvi Tummapudi, Caelan Alexander-Nordstrom, Daniel Brady, Eddy Nahmias. “AI-Assisted Medical Triage: Mitigating Performance Errors in Human-AI Emergency Response Training.” Presentation, Proc. Human Factors and Ergonomics Society 2026 Annual Meeting, Baltimore, MD, USA, 2026.

contact.md

Best way to reach me is through email or on GitHub.

colophon

This site is built with Astro ↗ and WebTUI ↗.