Senior Systems & Machine Learning Engineer designing large-scale personalization, search, recommendation, and GenAI systems.

• 🇮🇳 Born & raised in Mumbai, India
• 📍 Currently based in Boston, MA, US
• 🧠 ML Systems • Recommenders • GenAI / Agentic Workflows
• ⚡ Focus: retrieval → ranking → serving → experimentation
• 📧 [[email protected]]

• 🔬 AI Research Portfolio (12 projects): https://github.com/DJ92/ai-research-portfolio
  • Applied research: Constitutional AI, Safety, Evaluation, RAG, Interpretability
  • Foundational: Transformers, Pre-training, RLHF/DPO, LoRA & Quantization
• 📘 Blog & Technical Writing: https://dj92.github.io/interview-notes
  • 5 deep dives: Transformer Architecture, Pre-training, RLHF vs DPO, PEFT, Long Context LLMs
• 🧠 ML System Design: https://github.com/DJ92/ml-system-design
  • News Feed Ranking system (complete design), 7 more outlined
• 🤖 GenAI Systems: https://github.com/DJ92/genai
  • 9 production patterns: RAG, Agents, Guardrails, Multimodal RAG, Cost Optimization
• 🛠 Applied ML: https://github.com/DJ92/applied-ml
  • 11 projects: Classical ML, RecSys, Time Series, NLP, RL Fundamentals

• AI Safety & Alignment: Constitutional AI, preference learning (RLHF/DPO), guardrails, red-teaming
• Foundational ML: Transformer architectures, pre-training techniques, parameter-efficient fine-tuning
• LLM Evaluation: Automated metrics, LLM-as-judge, faithfulness analysis, interpretability
• Production GenAI: RAG systems, agentic workflows, tool use, prompt engineering
• ML Serving Systems: Low-latency inference, ranking, retrieval at scale
• Experimentation: Bridging offline metrics ↔ online impact via A/B testing
• ML Infrastructure: Feature stores, embeddings, vector search, training pipelines

Demonstrating both applied AI research and deep foundational knowledge → https://github.com/DJ92/ai-research-portfolio

Foundational Knowledge (Projects 9-12):

• Transformer Architecture from Scratch - Self-attention, multi-head attention, positional encodings (sinusoidal, learned, RoPE)

  • 38 tests, 74% coverage, full GPT implementation with generation

• Pre-training Techniques - Causal Language Modeling (CLM) & Masked Language Modeling (MLM)

  • 11 tests, 87% coverage, 80/10/10 masking strategy

• Post-Training Methods - Supervised Fine-Tuning, RLHF (reward model + PPO), DPO

  • 12 tests, 86% coverage, Bradley-Terry preference learning

• Parameter-Efficient Fine-Tuning - LoRA implementation, 4-bit/8-bit quantization

  • 8 tests, 81% coverage, 99% parameter reduction, 4-8× compression

Alignment & Safety Research (Projects 5-7):

• Constitutional AI & Preference Learning - Anthropic's critique-revision loop, RLHF simulation

  • 92% test coverage, 112% improvement in harmlessness, 82% preference accuracy

• CoT Faithfulness Analysis - Testing if reasoning actually drives answers via counterfactual interventions

  • 91% test coverage, 94% precision detecting unfaithful reasoning

• Agent Safety & Guardrails - Production safety mechanisms, prompt injection detection

  • 94% precision, 88% recall on injection detection, 73% attack reduction

Evaluation & Interpretability (Projects 1, 8):

• LLM Evaluation Framework - Automated metrics + LLM-as-judge, cost tracking

  • 87% coverage, 0.82 judge-human correlation

• Model Interpretability Toolkit - Attention analysis, logit attribution, perplexity metrics

  • r = -0.78 PPL-accuracy correlation, 82% attribution precision

Production Systems (Projects 2-4):

• Production RAG System - Chunking strategies, retrieval metrics, hybrid search

  • 85% MRR@10 with semantic chunking

• Tool Use & Function Calling - ReAct agents, schema validation

  • 94% tool selection accuracy

• Prompt Engineering Lab - Systematic evaluation of techniques

  • Few-shot optimal: 3-5 examples (91% accuracy)

5 Deep Dives on LLM Foundations & Production → https://dj92.github.io/interview-notes

• Transformer Architecture - Self-attention mechanics, positional encodings (sinusoidal/RoPE), O(n²) complexity analysis
• LLM Pre-training - CLM vs MLM, 80/10/10 masking, warmup+cosine decay, Chinchilla scaling laws
• Alignment: RLHF vs DPO - Reward modeling, PPO, Direct Preference Optimization, Constitutional AI
• Parameter-Efficient Fine-Tuning - LoRA (99% param reduction), 4-bit/8-bit quantization, QLoRA
• Long Context LLMs - RAG vs Long Context vs Hybrid, "lost in the middle" problem, cost-quality tradeoffs

Complete Architecture Designs → https://github.com/DJ92/ml-system-design

• News Feed Ranking (COMPLETE) - Two-stage retrieval+ranking, multi-task DNN, <100ms P99, 100M+ DAU scale
• 8 System Designs Outlined - Video recommendations, search ranking, RTB ads, fraud detection, email ranking, explore/exploit, embeddings

Production Patterns → https://github.com/DJ92/genai

• 9 Project Implementations - Production RAG (hybrid search, reranking), agentic code review (ReAct pattern), multi-agent support, prompt optimization, guardrails, semantic code search, document intelligence, cost optimizer (65% reduction), multimodal RAG (88% accuracy on text+image)

Classical to Modern Techniques → https://github.com/DJ92/applied-ml

• 11 Projects - Logistic regression from scratch, decision trees/random forests, gradient boosting, collaborative filtering, neural CF, sequential recommendations (GRU4Rec), ARIMA/LSTM time series, text classification (BERT), NER (BiLSTM-CRF), RL fundamentals (Q-learning → PPO, foundation for RLHF)

• mb-str - Multi-Behavior Sequential Transformer Recommender → https://huggingface.co/Djosh1992/mb-str

• Bias for action, with mechanisms to course-correct quickly
• Treat ML models as production software, not research artifacts
• Prefer simple, well-scoped designs with explicit trade-offs and failure modes
• Make decisions with imperfect data, then validate via experiments and metrics
• Care deeply about end-to-end ownership: data → model → serving → impact
• Balance model quality, latency, and reliability based on user and business goals
• Document assumptions, measure outcomes, and learn in public (within teams)
• Optimize for observability, debuggability, and iteration speed over premature complexity

My work follows a structured progression from foundations to production systems:

┌─────────────────────────────────────────────────────────────────┐
│                    FOUNDATIONS (12 projects)                     │
│  Transformer Architecture • Pre-training • RLHF/DPO • LoRA      │
│              ai-research-portfolio/09-12                         │
└────────────────────────┬────────────────────────────────────────┘
                         │
                         ▼
┌─────────────────────────────────────────────────────────────────┐
│              RESEARCH & SAFETY (8 projects)                      │
│  Constitutional AI • CoT Faithfulness • Agent Safety             │
│  LLM Evaluation • Interpretability • RAG • Tool Use              │
│              ai-research-portfolio/01-08                         │
└────────────────────────┬────────────────────────────────────────┘
                         │
                         ▼
┌─────────────────────────────────────────────────────────────────┐
│           PRODUCTION GenAI (9 projects + 5 blog posts)           │
│  Production RAG • Agents • Guardrails • Multimodal RAG           │
│  Cost Optimization • Long Context Strategies                     │
│         genai/ + interview-notes/                                │
└────────────────────────┬────────────────────────────────────────┘
                         │
                         ▼
┌─────────────────────────────────────────────────────────────────┐
│          ML SYSTEMS (1 complete + 7 designs, 11 projects)        │
│  News Feed Ranking • RecSys • Time Series • Classical ML • RL    │
│         ml-system-design/ + applied-ml/                          │
└─────────────────────────────────────────────────────────────────┘

Flow: Foundation → Research → Production → Systems demonstrates both depth (can build transformers from scratch) and breadth (can deploy at scale).

Research papers implemented across the portfolio:

• ✅ Attention Is All You Need - Transformer architecture (Project 09)
• ✅ RoFormer: Enhanced Transformer with Rotary Position Embedding - RoPE implementation (Project 09)

• ✅ BERT: Pre-training of Deep Bidirectional Transformers - MLM objective (Project 10)
• ✅ Language Models are Unsupervised Multitask Learners - GPT-2 CLM (Project 10)
• ✅ Training Compute-Optimal Large Language Models - Chinchilla scaling laws (Project 10)

• ✅ Training language models to follow instructions with human feedback - InstructGPT/RLHF (Project 11)
• ✅ Direct Preference Optimization - DPO implementation (Project 11)
• ✅ Constitutional AI: Harmlessness from AI Feedback - CAI critique-revision (Project 06)
• ✅ Measuring Faithfulness in Chain-of-Thought Reasoning - CoT faithfulness (Project 07)

• ✅ LoRA: Low-Rank Adaptation of Large Language Models - LoRA implementation (Project 12)
• ✅ QLoRA: Efficient Finetuning of Quantized LLMs - 4-bit quantization (Project 12)

• ✅ Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks - RAG system (Project 03, GenAI Project 01)
• ✅ Lost in the Middle: How Language Models Use Long Contexts - Long context analysis (Blog post)

• ✅ ReAct: Synergizing Reasoning and Acting in Language Models - ReAct agents (Project 04, GenAI Project 02)
• ✅ Chain-of-Thought Prompting Elicits Reasoning in Large Language Models - CoT prompting (Project 05)

• ✅ Neural Collaborative Filtering - NCF implementation (Applied ML Project 05)
• ✅ Session-based Recommendations with Recurrent Neural Networks - GRU4Rec (Applied ML Project 06)

• ✅ Proximal Policy Optimization Algorithms - PPO implementation (Applied ML Project 11)
• ✅ Playing Atari with Deep Reinforcement Learning - DQN (Applied ML Project 11)

Total: 22 papers spanning transformers, alignment, efficiency, RAG, agents, and RL.

Quantitative overview of work across all repositories:

• Lines of Code: 6,500+ (excluding tests and configs)
• Test Coverage: 85%+ average across AI research projects
• Documentation: 12,000+ lines across READMEs, blog posts, system designs
• Total Projects: 33 (12 AI research + 9 GenAI + 11 Applied ML + 1 complete system design)

• Papers Implemented: 22 foundational papers
• Quantitative Metrics: Every project includes measurable results (accuracy, latency, cost, coverage)
• Reproducibility: Fixed seeds, documented hyperparameters, requirements files
• Test Coverage by Project:
  • Constitutional AI: 92%
  • CoT Faithfulness: 91%
  • LLM Evaluation: 87%
  • Pre-training: 87%
  • Post-training: 86%
  • PEFT: 81%
  • Transformer Architecture: 74%

• System Designs: 1 complete (News Feed Ranking), 7 outlined
• Latency Optimizations: <100ms P99 (News Feed), <500ms P99 (RAG), <50ms (Guardrails)
• Cost Optimizations: 65% reduction (GenAI cost optimizer), 100× cheaper (RAG vs Long Context)
• Scale Targets: 100M+ DAU (News Feed), 10M ratings (RecSys), 1M tokens (Long Context)

• AI Safety: 3 projects (Constitutional AI, CoT Faithfulness, Agent Safety)
• Foundational ML: 4 projects (Transformers, Pre-training, RLHF/DPO, LoRA)
• Production GenAI: 9 projects (RAG, Agents, Guardrails, Multimodal, etc.)
• Classical ML: 11 projects (Logistic Regression → RL Fundamentals)
• Technical Writing: 5 deep-dive blog posts (2,000+ lines)

• Python: Primary (PyTorch, Transformers, LangChain, scikit-learn)
• Infrastructure: Docker, Kubernetes, GCP, AWS
• Monitoring: MLflow, wandb, custom evaluation frameworks
• Serving: FastAPI, TorchServe, model optimization

Portfolio demonstrates: Foundational depth (can build transformers) + Applied research (Constitutional AI, CoT) + Production systems (RAG, ranking, serving at scale).