Skip to content
View TahaUser5's full-sized avatar
4 followers · 4 following

Achievements

Achievement: Pull SharkAchievement: QuickdrawAchievement: YOLO

Achievements

Achievement: Pull SharkAchievement: QuickdrawAchievement: YOLO

Highlights

  • Pro

Block or report TahaUser5

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
TahaUser5/README.md 
╔══════════════════════════════════════════════════════════════╗
║                                                              ║
║         Building AI systems that work in production.        ║
║                                                              ║
╚══════════════════════════════════════════════════════════════╝

Taha Tanvir

AI Engineer · Generative AI · Machine Learning · Full Stack

MPhil Artificial Intelligence @ PUCIT, University of the Punjab

Portfolio LinkedIn Email Location Profile Views Years Badge Repos Badge

> whoami

I work across the full AI stack — from training deep learning models and fine-tuning diffusion systems, to building production RAG pipelines and shipping full-stack applications. RAG is my deepest speciality but not my only gear.

Currently completing my MPhil in AI at PUCIT. I care about one thing: AI that actually works when you deploy it — evaluated, observable, and production-ready.

Current focus:  Production RAG · Generative AI · Multimodal Systems
Background:     Full Stack Engineering (MERN · FastAPI · Flutter)  
Research:       Published HAR paper · CNN-LSTM multimodal fusion
Next stop:      AI Engineer roles — Lahore → Dubai → Ireland

> tech_stack --list

AI & Machine Learning

PyTorch Scikit-learn HuggingFace CNN LSTM GRU ViT Swin Deep Learning

Generative AI

Diffusers DreamBooth Stable Diffusion CLIP LoRA TTS

RAG & Information Retrieval

LangChain Pinecone Cohere FAISS BM25 RAGAS Groq

Backend & Infrastructure

FastAPI Flask Docker PostgreSQL MongoDB Firebase REST APIs

Web & Mobile

Next.js React Node.js Flutter TypeScript Tailwind

Languages

Python JavaScript C++ SQL

> ls projects/

🔍 RAG Knowledge Base System

Hybrid retrieval combining dense vector search (HuggingFace + Pinecone) and BM25, with Cohere reranking and Groq generation. Plug-and-play retriever/LLM architecture deployed via FastAPI + Docker + Streamlit.

Faithfulness:       1.0   ████████████
Context Precision:  ~1.0  ████████████

Python LangChain Pinecone Cohere Groq FastAPI Docker

🔒 Private

🎨 DreamBooth LoRA — SDXL

Fine-tuned SDXL (6.6B params) with LoRA Rank-32 on 5 photos per subject using 13 memory optimization techniques on a 15GB GPU. Full rembg + prior preservation pipeline.

CLIP-I Fidelity:  70.88%  ████████▏
SD 1.5 Baseline:  48.20%  █████▊
Improvement:      +47%    relative

PyTorch Diffusers SDXL LoRA CLIP rembg

Repo

📊 CIFAR-100 Architecture Benchmark

Systematic benchmarking of ResNet50, ViT-B/16, and Swin Transformer Tiny on CIFAR-100 using ImageNet pre-trained weights. Analyzed tradeoffs in convergence speed, memory, and generalization.

ResNet50:   82.22%  ████████▏
ViT-B/16:  87.73%  ████████▊
Swin Tiny: 87.23%  ████████▊  ← most efficient

Python PyTorch HuggingFace timm Transfer Learning

Repo

🏃 Multimodal HAR — Published Research

Early vs Late Fusion CNN-LSTM for 12-class activity recognition across smartphone, smartwatch, and smart glasses sensor streams. LOSO subject-independent validation on CogAge dataset.

Late Fusion Accuracy:  55.18%
Validation:            LOSO (subject-independent)
Status:                Published research paper

Python PyTorch CNN-LSTM CogAge Sensor Fusion

Paper

🗣️ AI Voice Cloning Application (FYP)

Pre-trained TTS deep learning models integrated via Flask backend for high-fidelity voice cloning from reference audio. Cross-platform Flutter mobile app with Firebase auth, audio storage, and real-time sync.

Python Flask Deep Learning TTS Flutter Firebase

Repo

🖼️ Image Captioning — Flickr30K

Comparative study of RNN baseline vs CLIP+GPT-2 vs BLIP on Flickr30K. Demonstrates the full evolution from statistical to multimodal approaches in image captioning.

BLIP CLIP GPT-2 InceptionV3 BLEU Flickr30K

Repo

> cat experience.log

2023 – 2024      Freelance Full Stack Developer · Fiverr (Remote)
                 Built full-stack data management system for Switzerland-based client
                 MERN stack · database architecture → React frontend → cloud deployment
                 Domain config · REST API integration · secure data handling

2025 Dec         IBM Full Stack Software Developer Professional Certificate
                 Coursera · Issued by IBM · Verified on Credly

2021 – 2025      BS Software Engineering
                 University of Lahore

2025 – Present   MPhil Artificial Intelligence
                 PUCIT, University of the Punjab · Lahore

> cat stats.json

GitHub Stats

Top Languages

Streak

> echo $QUOTE

"The question of whether a computer can think is no more interesting
 than the question of whether a submarine can swim."
                                         — Edsger W. Dijkstra

Currently open to AI Engineer and Full Stack roles in Lahore and Dubai.

Portfolio

Pinned Loading

  1. cifar100-architecture-benchmark cifar100-architecture-benchmark Public

    Benchmarking ResNet50, ViT-B/16, and Swin Transformer Tiny on CIFAR-100 image classification using PyTorch and transfer learning.

    Jupyter Notebook

  2. dreambooth-lora-sdxl dreambooth-lora-sdxl Public

    Few-shot subject-driven image generation using DreamBooth LoRA on SDXL — 5 photos, 15GB GPU, 70.88% CLIP-I fidelity

    Jupyter Notebook

  3. image-captioning-flickr30k image-captioning-flickr30k Public

    Image captioning model using InceptionV3 + LSTM trained on Flickr30k dataset — generates natural language descriptions for images with BLEU-1 evaluation.

    Jupyter Notebook

  4. Job_Board Job_Board Public

    A full-stack job board app with a Python web scraper, Node.js/Express backend, and React frontend for browsing and managing scraped job postings.

    Python

  5. self-attention-evolution-har self-attention-evolution-har Public

    Visualizing the evolution of self-attention mechanisms on multimodal HAR sensor data from the CogAge dataset.

    Jupyter Notebook

  6. SIR-Model-Verification-Validation SIR-Model-Verification-Validation Public

    SIR epidemic model with formal property verification, exhaustive parameter tuning, and prediction on Influenza data (USA & Netherlands, 2009-2011). Computational Modelling

    Python