Muhammad Idrees code-with-idrees

📌 Abstract

Undergraduate AI/ML researcher working across deep learning, high-performance computing, and generative AI — from CUDA-level kernel optimization to multi-agent LLM pipelines. My work is grounded in one question: how do we make models that are efficient, interpretable, and genuinely deployable?

researcher = {
    "name"        : "Muhammad Idrees",
    "affiliation" : "FAST-NUCES Islamabad — B.S. Computer Science",
    "location"    : "Rawalpindi, Pakistan 🇵🇰",
    "focus"       : ["Machine Learning", "HPC / CUDA", "Generative AI", "NLP"],
    "mission"     : "Build AI systems that are efficient, scalable, and universally accessible",
    "status"      : "Open to Research Collaborations · PhD Opportunities · Internships",
    "contact"     : "muhammad.idrees2k25@gmail.com",
}

📊 Impact Snapshot

🏆 Global Programs	📄 Research Output	🧪 Flagship Result	🌍 Reach
Stanford CiP · CERN BL4S	10-page LNCS-format report	24.62 dB PSNR on CIFAR-10 DAE	GCP × MLB Hackathon
2 international programs	45+ ablation experiments	0.8225 SSIM · 182K params	6+ shipped projects

🧭 Research Focus Map

graph TD
    A["🧑‍💻 Muhammad Idrees"] --> B["🧠 Machine Learning & NLP"]
    A --> C["🖥️ High-Performance Computing"]
    A --> D["🤖 Generative AI"]
    A --> E["🔬 Emerging Directions"]

    B --> B1[Transformers: BERT · T5]
    B --> B2[Seq2Seq & Low-Resource NMT]
    B --> B3[Multi-Agent LLM Orchestration]
    B --> B4[LoRA · QLoRA · PEFT]

    C --> C1[CUDA Kernel Programming]
    C --> C2[MPI / OpenMP Parallelism]
    C --> C3[FlashAttention & Efficient Inference]
    C --> C4[Model Compression & Quantization]

    D --> D1[Gemini · Groq · Ollama]
    D --> D2[Diffusion & Image Synthesis]
    D --> D3[Autonomous Tool-Use Agents]
    D --> D4[Prompt Engineering & RLHF]

    E --> E1[Sparse Mixture-of-Experts]
    E --> E2[Memory-Efficient Transformers]
    E --> E3[Multi-Agent Reasoning Systems]
    E --> E4[Foundation Model Pre-training]

    style A fill:#0d1117,stroke:#64b5f6,stroke-width:2px,color:#e2e8f0
    style B fill:#0a1628,stroke:#64b5f6,color:#e2e8f0
    style C fill:#0a1628,stroke:#64b5f6,color:#e2e8f0
    style D fill:#0a1628,stroke:#64b5f6,color:#e2e8f0
    style E fill:#0a1628,stroke:#64b5f6,color:#e2e8f0

🗓️ Timeline

timeline
    title Research & Achievement Timeline
    2025 : Stanford Code in Place — Section Leader
         : CERN Beamline for Schools (BL4S)
    2025 : Google Cloud × MLB Hackathon
         : LNCS Research Report — CIFAR-10 Denoising Autoencoder
    Ongoing : Technical writing on Medium
            : Open-source project development

🔬 Research & Interests

🧠 Machine Learning & NLP

Transformer architectures (BERT, T5, encoder-decoder)
Sequence-to-sequence & low-resource NMT
Multi-agent LLM orchestration
Efficient fine-tuning: LoRA, QLoRA, PEFT

🖥️ High-Performance Computing

CUDA programming & GPU-accelerated ML
Parallel & distributed training (MPI, OpenMP)
Memory-efficient inference & FlashAttention
Model compression & quantization

🤖 Generative AI

Large language models (Gemini, Groq, Ollama)
Diffusion models & image synthesis
Autonomous AI agents & tool-use pipelines
Prompt engineering & RLHF

🧭 Currently Exploring

Sparse Mixture-of-Experts architectures
FlashAttention & memory-efficient transformers
Autonomous multi-agent reasoning systems
Foundation model pre-training at scale

🎯 Expertise Matrix

quadrantChart
    title Depth vs. Breadth Across Focus Areas
    x-axis "Exploring" --> "Core Strength"
    y-axis "Applied" --> "Research-Grade"
    quadrant-1 "Deep Research Focus"
    quadrant-2 "Applied Depth"
    quadrant-3 "Early Exploration"
    quadrant-4 "Broad Applied Skill"
    "Deep Learning / PyTorch": [0.85, 0.8]
    "NLP / Transformers": [0.8, 0.75]
    "CUDA / HPC": [0.65, 0.7]
    "Generative AI / Agents": [0.75, 0.55]
    "Cloud / MLOps": [0.55, 0.4]
    "MoE / FlashAttention": [0.4, 0.85]

_{Self-assessed positioning across active focus areas — not a benchmarked metric.}

🥇 Featured Research

🖼️ Denoising Autoencoder for CIFAR-10 — LNCS-Format Research Report

PyTorch Deep Learning Signal Reconstruction

Problem. Reconstruct clean images from noisy CIFAR-10 inputs while keeping the model small enough for constrained deployment.

Approach. Designed a lightweight, fully custom denoising autoencoder and ran a systematic ablation sweep across architecture depth, noise schedules, and latent bottleneck sizing — 45+ configurations in total — writing the full study up as a 10-page LNCS-format research report.

Result.

Metric	Value
Parameters	182K
PSNR	24.62 dB
SSIM	0.8225
Ablation experiments	45+
Report format	10-page, LNCS

xychart-beta
    title "Illustrative Ablation Trend — PSNR vs. Model Capacity"
    x-axis ["32K", "64K", "96K", "128K", "160K", "182K"]
    y-axis "PSNR (dB)" 18 --> 26
    line [19.8, 21.4, 22.6, 23.5, 24.1, 24.62]

Chart illustrates the general capacity–quality trend observed across the ablation sweep; see the full report for the complete 45+ experiment breakdown.

💡 Selected Projects

📖 Intelligent RC & Quiz Generation

PyTorch BERT T5 NLP

Problem: Automate reading-comprehension question generation from raw text. Approach: Hybrid pipeline pairing BERT for comprehension/context encoding with T5 for question generation. Outcome: A full AI-powered educational tool for auto-generating quizzes.

🖼️ Denoising Autoencoder — CIFAR-10

PyTorch Deep Learning Research

Problem: Reconstruct clean signal from noisy images at low parameter cost. Approach: 182K-parameter DAE, 45+ ablation experiments. Outcome: 24.62 dB PSNR / 0.8225 SSIM, documented in a 10-page LNCS report.

🎬 Multi-Agent AI Video Generation

Python Gemini API Groq LLMs

Problem: Automate end-to-end AI video creation from a prompt. Approach: Modular multi-agent pipeline — specialized LLM agents for scriptwriting, scene planning, and synthesis, with Gemini primary and Groq fallback for resilience. Outcome: A working automated video generation pipeline.

🤝 Google Meet AI Attendance Agent

Faster-Whisper Ollama Playwright

Problem: Automate meeting attendance, note-taking, and Q&A without cloud APIs. Approach: Offline agent using Playwright to join meetings, Faster-Whisper for transcription, and Ollama for local LLM reasoning. Outcome: Fully offline attendance tracking + auto-generated structured PDF notes, zero API keys.

🌐 English–Urdu Neural Machine Translation

PyTorch RNN Seq2Seq NLP

Problem: Translate a genuinely low-resource language pair (English ↔ Urdu). Approach: Vanilla RNN encoder-decoder, with an empirical study of vanishing gradients, sequence bottlenecks, and training instability under low-resource conditions. Outcome: A working NMT baseline plus a documented failure-mode analysis.

⚾ Google Cloud × MLB Hackathon

GCP Python Predictive Modeling

Problem: Improve live fan engagement using real-time data. Approach: Production-grade platform on Google Cloud AI with real-time predictive models and personalized content pipelines. Outcome: A deployed fan-engagement platform built for a live hackathon.

🛠️ Technical Stack

Languages

ML / Deep Learning

Generative AI & LLMs

HPC & Systems

Cloud & MLOps

🏅 Highlights


🎓 Stanford Code in Place — Section Leader `2025`	Selected from a global pool to mentor students in Python & computational thinking for Stanford's international CS program
🔬 CERN Beamline for Schools (BL4S) `2025`	Competed in CERN's international physics competition, proposing an original experiment applying HPC and ML concepts
☁️ Google Cloud × MLB Hackathon `2025`	Built a production-grade AI fan engagement platform on GCP with real-time predictive models
📄 LNCS Research Report — Denoising Autoencoder `2025`	Produced a full 10-page academic paper with 45+ ablation figures on CIFAR-10 image denoising
✍️ AI Research Writing on Medium	Publishing technical deep-dives on ML, NLP, and systems for a global audience

📈 GitHub Activity

✍️ Writing

Research articles, tutorials, and ML deep-dives on Medium →

🤝 Open For

🔬 Research collaborations in NLP, HPC, or generative AI
🎓 PhD / graduate research opportunities
💼 ML/AI internships
🗣️ Speaking, mentoring, or reviewing for student research programs

Machine Learning · HPC · Generative AI

Rawalpindi, Pakistan → Open to Remote Research & PhD Opportunities Worldwide

Navigation Menu