• AI Engineer working with deep learning, speech systems, and large language models.
• Currently pursuing Data Science at Air University Islamabad, in my 7th semester.
• Started working professionally in AI during my 5th semester.
• I work across model architecture, training, fine-tuning, optimization, inference, and deployment.
• Strong focus on Speech-to-Text, Text-to-Speech, speech translation, and low-resource languages.
• Worked extensively with Pakistani Pashto and challenging real-world speech data.
• Experienced with LLM training, fine-tuning, embeddings, synthetic data generation, and model adaptation.
• Build production AI systems using FastAPI, Docker, Kafka, CUDA, and GPU inference pipelines.
• Interested in model optimization, inference engines, Apple Silicon, GPU computing, and open-source AI infrastructure.
• Also involved in mentoring junior engineers and interns in machine learning and deep learning.

• Automatic Speech Recognition
• Text-to-Speech
• Speech-to-Speech Translation
• Low-resource language modeling
• Whisper fine-tuning and optimization
• Voice and acoustic modeling
• Neural audio pipelines
• Streaming and real-time inference

• LLM fine-tuning
• Parameter-efficient training
• LoRA and DoRa
• Synthetic data generation
• Model evaluation
• Quantization
• Local LLM deployment
• Agentic AI systems

• GPU inference optimization
• CUDA workloads
• Quantized inference
• CTranslate2
• Apple Silicon / MPS
• Memory optimization
• Latency optimization
• Production inference pipelines

• FastAPI
• Docker
• Kafka
• REST APIs
• Model serving
• GPU deployments
• Data pipelines
• Git and GitHub

Built custom speech pipelines involving technologies such as:

Whisper → Neural Adapters → G2P / Linguistic Processing → TTS

Worked on speech recognition and translation under heavily degraded GSM and telephony audio, where noise, compression, low bandwidth, and unclear speech make conventional models struggle.

My work includes:

• Custom Whisper fine-tuning
• Encoder adaptation using LoRA
• Decoder hidden-state extraction
• Neural representation adapters
• Transformer and CTC architectures
• Temporal convolution bridges
• NeMo-based linguistic processing
• VITS-based speech synthesis
• End-to-end GPU inference pipelines

Developed an English-to-Pashto speech-to-speech research pipeline combining:

Whisper + Neural Adapters + NeMo G2P + Pashto VITS

Evaluation on an RTX 3090:

Generated Audio: 11.006 seconds
End-to-End Inference: 4.436 seconds
Real-Time Factor: 0.403

The complete pipeline operates significantly faster than real time.

Working on an experimental Apple Silicon MPS backend contribution for OpenNMT/CTranslate2.

The work involves extending a high-performance inference library beyond its traditional CPU and CUDA execution paths and evaluating:

• MPS operator support
• Model correctness
• WER / BLEU preservation
• Performance
• Memory behavior
• Apple Silicon compatibility

This work involves lower-level model execution and inference infrastructure rather than treating ML frameworks as black boxes.

Built Vibify, an AI-powered music recommendation platform using:

FastAPI • OpenAI Embeddings • Pinecone • Python

Implemented semantic music recommendation using vector embeddings and similarity search.

interests = [
    "Speech AI",
    "Large Language Models",
    "Low-Resource Languages",
    "Real-Time AI",
    "Model Optimization",
    "Quantization",
    "Inference Engines",
    "Apple Silicon AI",
    "GPU Computing",
    "Open Source AI"
]

I particularly enjoy understanding how models work internally, modifying architectures, optimizing inference paths, and turning research models into production systems.

Building AI systems from model architecture to production inference.