Portfolio Content Overview

📁 File Structure

/Users/arjun/Documents/portfolio-site/
│
├── index.md                          # 🏠 Homepage (5.8 KB)
├── _config.yml                       # ⚙️ Jekyll configuration
├── README.md                         # 📖 Setup instructions
├── DEPLOYMENT_CHECKLIST.md           # ✅ Deployment guide
├── CONTENT_OVERVIEW.md              # 📋 This file
│
└── case-studies/
    ├── kawa-space.md                # 📡 Satellite imagery (11 KB)
    ├── truecaller.md                # 🔒 Fraud detection (17 KB)
    └── gpt2-from-scratch.md         # 🤖 GPT-2 tutorial (24 KB)

Total Content: ~58 KB of high-quality technical writing

📄 Page-by-Page Breakdown

🏠 Homepage (index.md)

Sections:

  1. Hero/Elevator Pitch - “I bridge the gap between research and production”
  2. Tech Stack - Python, Scala, PyTorch, MLFlow, GCP/AWS, etc.
  3. Featured Work - 3 cards with summaries + links to case studies
  4. Competitions - Microsoft AI, IEEE Fraud, LANL Earthquake
  5. Advisory/Consulting - GenAI, VoiceTech projects
  6. Experience Timeline - All roles with key impacts
  7. Education - Jadavpur University
  8. Contact - Email, LinkedIn, GitHub

Key Metrics Highlighted:

  • 10 years experience
  • 25% cost reduction
  • 8% fraud detection improvement
  • 5x latency improvement
  • $10K contract landed

Tone: Authoritative, impact-driven, senior engineer

📡 Kawa Space Case Study (case-studies/kawa-space.md)

Sections:

  1. Executive Summary - Quick overview with key impact
  2. The Problem - 4 technical challenges explained
  3. The Solution - Multi-pronged approach:
    • Data standardization algorithm (with code samples)
    • End-to-end GCP pipeline (architecture diagram)
    • Land boundary detection model (U-Net)
    • Performance benchmarking
  4. The Impact - Business + technical achievements
  5. Technical Deep Dive - Implementation details
  6. Conclusion - Key learnings

Key Metrics:

  • 5x latency improvement (150s → 30s)
  • $10K+ first contract
  • 35% net cost reduction
  • 0.87 IoU accuracy

Code Samples: Python pseudocode for:

  • Spatial standardization
  • Radiometric normalization
  • Metadata handling
  • TFRecord creation

What It Shows:

  • ✅ System architecture skills
  • ✅ GCP expertise (Cloud Functions, Dataflow, AI Platform)
  • ✅ Computer vision (semantic segmentation)
  • ✅ Business impact orientation

🔒 Truecaller Case Study (case-studies/truecaller.md)

Sections:

  1. Executive Summary - Billion-row scale, cost optimization
  2. The Problem - 4 major challenges
  3. The Solution - 3 innovations:
    • Automated retraining with drift detection (MLFlow)
    • PySpark optimization (6 strategies)
    • Semi-supervised learning (DeBERTaV3)
  4. The Impact - Cost + performance improvements
  5. Technical Deep Dives - Detailed implementations
  6. Conclusion - “Production ML is 80% infrastructure”

Key Metrics:

  • 25% cloud cost reduction
  • 8% fraud detection improvement
  • 10% SMS spam detection improvement
  • Billions of rows processed
  • 99.9% uptime

Code Samples: Complete implementations of:

  • Data drift detector (KS test, KL divergence)
  • MLFlow pipeline integration
  • Spark optimization techniques
  • Semi-supervised learning with pseudo-labeling

What It Shows:

  • ✅ MLOps mastery
  • ✅ Big data engineering (PySpark)
  • ✅ Cost optimization
  • ✅ NLP (transformers, DeBERTa)
  • ✅ Real-world scale (billions of rows)

🤖 GPT-2 Tutorial (case-studies/gpt2-from-scratch.md)

Sections:

  1. Overview - What you’ll learn
  2. Architecture Explanation - High-level understanding
  3. Multi-Head Attention - Math + complete PyTorch code
  4. Transformer Block - Feed-forward + residual connections
  5. Full GPT-2 Model - 117M parameter implementation
  6. Training Pipeline - Dataset prep, training loop, optimization
  7. Text Generation - Greedy, top-k, top-p sampling
  8. Results - Training metrics, sample generations
  9. Lessons Learned - What worked, challenges, improvements

Key Metrics:

  • 117M parameters (exact match to GPT-2 Small)
  • Perplexity: 30.6 (close to OpenAI’s 29.4)
  • 18 hours training on A100 GPU

Code Samples: Production-ready PyTorch:

  • Complete multi-head attention (with causal masking)
  • Transformer block (pre-norm architecture)
  • Full GPT-2 model (with weight tying)
  • Training loop (AdamW, lr scheduling, gradient clipping)
  • 3 sampling strategies (greedy, top-k, top-p)

What It Shows:

  • ✅ Deep theoretical knowledge
  • ✅ PyTorch expertise
  • ✅ Transformer architecture mastery
  • ✅ Training pipeline engineering
  • ✅ Teaching ability (tutorial format)
  • ✅ First-principles thinking

🎯 Strategic Positioning

Your Unique Value Proposition

“Senior ML Engineer who delivers production-grade, cost-optimized systems with measurable business impact”

What Differentiates This Portfolio

  1. Depth over Breadth
    • 3 deep-dive case studies vs. 20 shallow project descriptions
    • Real production metrics (25% cost reduction, 5x latency)
    • Actual business impact ($10K contract, 8% fraud improvement)
  2. Production Focus
    • Every project shows end-to-end systems, not just models
    • MLOps, cost optimization, scalability are central
    • Real-world constraints (billion-row datasets, cloud costs)
  3. Technical Rigor
    • Code samples throughout (not just descriptions)
    • Architectural diagrams and performance benchmarks
    • Honest trade-offs and learnings sections
  4. Business Acumen
    • Links technical work to business outcomes
    • Shows understanding of ROI, SLAs, customer needs
    • Demonstrates senior-level strategic thinking

Target Audience

Primary: Hiring managers for senior ML/AI roles at:

  • Tech companies (Google, Meta, Amazon, Microsoft)
  • Fast-growing startups (Series B+)
  • Fintech, SaaS, E-commerce companies

Secondary:

  • Recruiters specializing in ML/AI
  • Potential consulting clients
  • Technical peers (for networking)

Key Messages by Case Study

Case Study Primary Message Secondary Message
Kawa Space I can architect end-to-end ML systems on cloud platforms I understand business impact (landed $10K contract)
Truecaller I optimize for cost and scale, not just accuracy I build production MLOps systems that run themselves
GPT-2 I understand ML deeply (first principles) I can teach and explain complex concepts clearly

📈 Next Steps: Content Evolution

Phase 2: Add Blog Posts (Future)

After deployment, consider adding these:

  1. “5 Lessons from Processing Billions of Rows with PySpark”
    • Extract from Truecaller case study
    • More tutorial-focused
    • SEO-friendly title
  2. “How We Reduced ML Infrastructure Costs by 25%”
    • Detailed cost breakdown
    • Before/after comparisons
    • Actionable tips for readers
  3. “From Research Paper to Production: Implementing DeBERTa for Spam Detection”
    • Bridge research → production gap
    • Show how to fine-tune transformers
    • Include code repository

Phase 3: Add Code Repositories

Create GitHub repos for:

  1. gpt2-from-scratch - Full implementation + training scripts
  2. spark-optimization-toolkit - Reusable Spark utilities
  3. mlflow-drift-detector - Open-source drift detection library

Link these from case studies to show “see my code in action”

Phase 4: Add Visualizations

Consider adding:

  • Architecture diagrams (draw.io or Excalidraw)
  • Performance charts (matplotlib/seaborn)
  • GIF demos of text generation
  • Screenshots of MLFlow experiments

🎨 Design Recommendations

Theme Options (in order of recommendation)

  1. jekyll-theme-minimal - Clean, professional, resume-like
    • Best for: Technical audience
    • Pros: No distractions, fast loading
    • Cons: Less visual appeal
  2. jekyll-theme-cayman - Modern with gradient header
    • Best for: Broader audience
    • Pros: Polished, professional
    • Cons: Slightly less code-focused feel
  3. minima (default) - Balanced, blog-style
    • Best for: General purpose
    • Pros: Well-maintained, feature-rich
    • Cons: Generic look

Recommendation: Start with minima, switch to minimal if you want more “senior engineer” vibe.

Typography Tips

  • Use ### for section headers (clear hierarchy)
  • Use bold for key metrics (25%, 5x, $10K)
  • Use code blocks for technical terms (MLFlow, PySpark)
  • Use > blockquotes for key insights
  • Use tables for comparisons (before/after, model sizes)

✅ Pre-Deployment Checklist

  • Homepage created with hero, tech stack, featured work
  • All 3 case studies written (Kawa, Truecaller, GPT-2)
  • Links between pages work (relative paths)
  • Code samples included throughout
  • Metrics and impact highlighted
  • GitHub Pages config file (_config.yml)
  • Setup instructions (README.md)
  • Deployment guide (DEPLOYMENT_CHECKLIST.md)

Missing (optional):

  • Resume PDF in resume/ folder
  • Profile photo (add to homepage)
  • Custom domain (if desired)
  • Google Analytics setup

🚀 Ready to Deploy!

Everything is ready. Follow the steps in DEPLOYMENT_CHECKLIST.md to go live in 5 minutes.

Your portfolio URL will be: https://Peaceout21.github.io

Questions or need changes? Let me know and I can update any content before deployment!